Does Importing Intermediates Increase the

Demand for Skilled Workers? Plant-level

Evidence from Indonesia∗

Hiroyuki Kasahara†

Vancouver School of Economics

University of British Columbia

Yawen Liang

Vancouver School of Economics

University of British Columbia

Joel Rodrigue

Department of Economics

Vanderbilt University

September 24, 2015

Abstract

This paper examines whether importing has contributed to skill upgrading among

Indonesian plants. Our data records the distribution of years of employee schooling

in each plant. We examine how importing affects the demand for highly educated

workers within both production and non-production occupation categories at the

plant level. We estimate a model of importing and skill-biased technological change

in which selection into importing arises due to unobservable heterogenous returns

from importing. Both instrumental variable regression and marginal treatment

effect estimates confirm that importing has substantially increased the relative de-

mand for educated production workers.

∗The authors thank the seminar participants at the Bank of Canada, UIBE, Purdue, VanderbiltUniversity, Rocky Mountain Empirical Trade Conference, Hitotsubashi COE Conference on Trade andFDI, Empirical Trade Conference at Georgia Institute of Technology and the Conference in Honor of BillEthier for their helpful comments. This research was supported by the SSHRC.†Address for correspondence: Hiroyuki Kasahara, Vancouver School of Economics, University of

British Columbia, 997-1873 East Mall, Vancouver, BC, V6T 1Z1 Canada.

1 Introduction

Workhorse models of international trade almost universally suggest that increased in-

tegration into international markets will encourage resources to be reallocated towards

workers, firms, or industries in which the country has a comparative advantage. In de-

veloping countries, for example, trade liberalization is often supported by the argument

that trade will expand in labor-intensive industries which, in turn, are predicted to in-

crease the relative demand and wages for unskilled labor. Surprisingly, in many contexts,

exactly the opposite has been found. Numerous studies confirm that among develop-

ing countries, trade liberalization has increased the relative firm-level demand for skilled

labour (Sanchez-Paramo and Schady, 2003; Goldberg and Pavcnik, 2007) and, likewise,

has caused the skill premium to rise (Harrison and Hanson (1999), Gindling and Robbins

(2001), Attanasio et al. (2004)).1 Despite these stark trends, the underlying cause of the

increased demand for skilled workers, the contribution from trade, and the implications

for income inequality remain key, unresolved issues (Goldberg and Pavcnik, 2005).2

This paper contributes to this literature by examining the impact that importing

foreign materials has on the demand for highly educated workers among Indonesian man-

ufacturing plants. The idea that importing may affect firm organization or productivity

is neither new or controversial. Rather, it is widely reported that using foreign interme-

diate goods in production often requires the plant-level adoption of more sophisticated

technology.3 The adoption of foreign technology, and thus importing in a developing

country, is likely to induce further structural changes within individual manufacturing

1In contrast, Amiti and Cameron (2012) find that falling input tariffs has caused the wage skillpremium within firms that import their intermediate inputs to fall.

2Our work is likewise related to studies of trade, employment and wages (Trefler, 2004; Gonsaga etal., 2006; Bernard et al, 2007; Egger and Kreikemeier, 2009; Davis and Harrigan, 2011; Felbermayr et al.,2011; Amiti and Davis, 2012), studies of trade, wages and the demand for skilled workers (Bernard andJensen, 1997; Yeaple, 2005; Verhoogen, 2008; Frıas et al., 2009; Chor, 2010; Helpman et al., 2010; Bustos,2011; Cosar, 2011; Vannoorenberghe, 2011), and studies of trade, wages and skill-biased technologicalchange (Feenstra and Hanson, 1999; Matsuyama, 2007; Costinot and Vogel, 2010; Bloom et al., 2011;Burstein and Vogel, 2012; Burstein et al., 2013; Parro, 2013).

3This is particularly true when it is imported from industrialized nations for which there is substantialevidence of skill-biased technological change. Doms, Dunne, and Troske (1997) provide evidence that theadoption of new factory automation technologies lead to skill upgrading. Further, recent literature ontrade and heterogeneous firms suggests that importing foreign intermediate goods increases productivity.See Muendler (2004), Amiti and Konings (2007), Kasahara and Rodrigue (2008), Halpern, Koren, andSzeidl (2015), and Kugler and Verhoogen (2009) among others. There is also significant evidence thatskill-biased technological change can increase the skill-premium even in developing countries (e.g., Kijima,2006). Burstein et al. (2013) provide an alternative model whereby importing directly induces skill-biasedtechnological change.

1

plants. In fact, there is a rich literature indicating that the reallocation of workers is

strongly related to changes in the demand for skilled labour within firms, rather than

across industries (Berman, Bound, and Griliches, 1994; Bernard and Jensen, 1997; and

Biscourp and Kramarz, 2007). We extend this line of research by relating changes in the

relative use of educated workers to observable decisions to import intermediate materials

at the plant-level.

Our data are exceptionally well suited to this objective. Typically, researchers have

used variation in occupation categories, such as non-production or white-collar workers,

to construct a proxy for skilled labor (Bernard and Jensen, 1997; Harrison and Hanson,

1999; Pavcnik, 2003; Biscourp and Kramarz, 2007).4 A major advantage of this paper’s

study is that it is able to capture a much more precise measure of skill at the plant-

level. Specifically, our panel data record the education-level of every worker in every

Indonesian manufacturing plant with at least 20 employees. Moreover, we are able to

distinguish the distribution of worker education across non-production and production

workers within each plant. This allows us to disentangle the impact of importing across

broad occupation categories and more precisely characterize the effect of importing on

firms in a developing country.

Quantifying the impact of importing on the demand for skilled labour requires over-

coming a number of key empirical challenges. First, we are particularly concerned that

the demand for skill and the decision to import are endogenously determined. We develop

a number of detailed instruments to capture variation in plant-level shipping import costs

and identify the causal impact of importing on the demand for skilled labour. Specifically,

we collect data on the shipping costs to major Indonesian ports, differences in the heavi-

ness of imports, and the fraction of imports shipped by air, along with product-level tariff

4Important exceptions are Bustos (2011) and Koren and Csillag (2011). Using a panel of Argentineanmanufacturing firms with the detailed information on worker’s education level Bustos (2011) finds thatexporting increases the demand for skilled labor, while our results suggest that importing, rather thanexporting, is more important for skill upgrading. Using Hungarian linked employer-employee data, Korenand Csillag (2011) find that the wage gap between workers with a high school diploma and those withprimary schooling is larger among workers operating imported machines than among workers operatingdomestic machines. Similarly, a number of studies use linked data firm and employee data to establisha number of related findings. In particular, Frazer (2013) examines the effect of importing on Rwandanmanufacturing wages, Hummels (2014) characterizes the relationship between offshoring and wages acrossskilled and unskilled Danish workers, Martins and Opromolla (2009) investigates the impact of importingon Portugese manufacturing wages, while Krishna et al (2011) and Menezes-Filho and Muendler (2011)study how trade reform affects Brazilian wages and worker displacement, respectively. Ebenstein et al(2014) examine the impact of offspring on US wages using CPS. While these studies offer insight intothe effect of importing on the workers’ wages or displacement, we examine the effect of importing on therelative employment of educated workers to less educated workers at the plant-level.

2

changes, to distinguish differences in the probability of importing across plants. Using

a variety of instrument sets we are able to capture robust IV estimates of the impact of

importing on the demand for skilled labor.

Second, we are also concerned that the impact of trade on the demand for educated

workers within plants will vary substantially across heterogeneous plants. For instance,

importing foreign intermediate goods may provide plants with an incentive to hire more

educated workers, but the degree of skill-upgrading may depend crucially on the plant’s

existing, potentially unobserved, heterogeneous ability to implement foreign technology.

When the effect of importing on the demand for skill varies across plants, there is no

single “effect” of importing on skill demand. Furthermore, we expect plants with greater

ability to adopt technology will self-select into importing, leading to a difficult selection

problem. In such cases, the instrumental variable (IV) estimator identifies the average

effect of importing among plants induced to change their import status by the instrument

(Imbens and Angrist, 1994). We also estimate the model by applying the treatment

effect framework developed by Heckman and Vytlacil (2005, 2007a, 2007b) to identify

various summary measures of the impact of importing on the relative demand for skilled

labour in the Indonesian manufacturing sector, such as the average effect among all plants

(the average treatment effect; the ATE, hereafter), the average effect among importers

(the treatment effect on the treated; the TT, hereafter), and the average effect among

non-importers (the treatment effect on the untreated; the TUT, hereafter).

Third, while we are able to identify the ATE, the TT and the TUT, it is unclear

whether these objects are of particular interest to the policymaker. For instance, while

the TT suggests that importing had an important impact on the demand for skilled labour

among plants that were induced to import in our sample, it is unclear that further policy

change will greatly affect the demand for skilled labour among new importers. We use

the estimation framework of Carneiro, Heckman, and Vytlacil (2010) to study the impact

of further policy changes on the demand for skilled labor among the set of plants induced

to import by the change in policy.

On one hand, we find that importing has had a large impact on the demand for

educated production workers among Indonesian manufacturing plants; across different

specifications, our IV estimates suggest that starting to import strongly increases the

demand for skilled labour among production workers. On the other hand, our IV estimates

do not provide robust evidence for any impact of importing on skill upgrading among non-

production workers. Our marginal treatment effect (MTE, hereafter) estimates not only

confirm that starting to import has strongly increased the demand for skilled production

3

workers but also indicate that not all plants benefit equally from importing. We find that

the ATE, the TT, and the TUT for production workers are significantly positive while

the TT is estimated to be substantially larger than the ATE, which, in turn, is estimated

to be larger than the TUT. Moreover, we find that further policy changes that promote

importing would have substantially increased the demand for skilled production workers

among the plants that would have been induced to start importing by policy changes.

The next section describes our data set and documents the relationship between im-

porting and plant-level skill-intensity. Section 3 describes our empirical model and the

nature of selection. Section 4 describes the empirical results. The last section concludes.

2 Data

2.1 Data Sources

Our primary source of data is the Indonesian manufacturing survey between 1995 and

2007, where we mainly use the data recorded in the census years 1996 and 2006 because, in

these two years, the Indonesian manufacturing survey records the distribution of academic

achievement in two distinct occupation categories (non-production vs. production) in

each plant. Specifically, in each plant we observe the number of workers with primary,

secondary and post-secondary education. We use this information to construct the relative

skill measures for each occupation category, which are directly based on the workers’

education levels.

The survey covers all manufacturing plants with at least 20 employees. In the 2006

data set, 93 percent of plants are also single-plant firms. The data set captures a wide set of

plant-level characteristics which we use to study the nature of plant-level heterogeneity.

In particular, the survey records all expenditures on imported intermediate materials

which we use to identify the effect of importing on the demand for skilled labor. It

also includes key plant-level input and output variables, such as total revenues, capital

stock, domestic materials, necessary for computing plant-level productivity. Likewise, the

survey records plant-level information including the percentage of sales from exports, the

percentage of ownership held by foreign investors, total plant-level expenses on research

and development (R&D), and total plant-level expenditures on worker training. Appendix

A provides a detailed description of our variable construction.

Naturally, the extent to which a plant is willing to hire skilled workers depends heavily

on the premium it must pay for skilled workers in its local labor market. Unfortunately,

4

the manufacturing survey data do not provide a measure of wages by education level.

To capture regional (city-level) variation in the skill premium we augment manufacturing

survey with the Indonesian household survey. The Indonesian household survey covers a

nationally representative sample of households. Each survey documents key labor force

information including gender, age, location, educational attainment and labor force ex-

perience among a wide set of additional characteristics. We use the household survey to

develop a measure of the skill premium in each location and year.

2.2 Importing and Worker Education

Panel A of Table 1 documents plant-level differences in employment across six education-

based (highest attainment) categories: less than primary school, primary school, junior

high school, high school, college graduates and post-graduate educated workers. The

top panel compares the percentage of plant-level employment across importing and non-

importing plants in 2006. We find that importing plants, on average, hire fewer workers

in each educational category below high school and more workers with high-school diplo-

mas, college degrees, or post-graduate training. For example, 61 percent of workers in

importing plants have at least a high school degree, while only 36 percent of workers

in non-importing plants have a high school degree or better. Columns (8)-(10) further

compare associated measures of skill-intensity across importing and non-importing plants.

“Training/Worker” and “R&D/Worker” report the average per worker expenditures on

training and research and development (R&D), respectively, in thousands of 1983 Indone-

sian rupiahs while “Non-Prod./All Workers” reports the percentage of non-production

workers in total employment in each plant. We find that the expenditures on training

workers or investing in R&D among importers is more than double what is spent by

non-importers on average. Likewise, importers tend to have a relatively large number of

non-production workers in their plants. Each of these measures strongly suggests that

importers are relatively skill-intensive manufacturing plants when compared to their non-

importing counterparts.

Panel A of Table 1 also compiles similar statistics for exporting plants, non-exporting

plants, domestic plants, and foreign-owned plants.5 We observe a number of stark pat-

terns: foreign plants tend to employ more skilled workers than domestic plants while

exporting plants appear skill-intensive when compared to their non-exporting counter-

parts. Nonetheless, within each group we continue to find that importing plants hire a

5We classify a plant as foreign plant when at least 10 percent of its equity is held by foreign investors.

5

Table 1: Importing and Skill Intensity 2006

Panel A: All WorkersHighest Degree Completed/Fraction

No Jr. Grad. Training R&D Non-Prod. No. of

Primary Primary High High College School Worker Worker All Workers Obs.All Plants

Importers 0.015 0.071 0.302 0.538 0.073 0.0006 70.9 73.8 0.184 5,512(0.075) (0.171) (0.218) (0.237) (0.093) (0.004) (724.4) (1037.1) (0.151)

Non-Importers 0.059 0.275 0.306 0.323 0.036 0.0003 23.2 17.8 0.135 23,952(0.151) (0.302) (0.248) (0.295) (0.078) (0.005) (570.0) (294.4) (0.163)

Exporting PlantsImporters 0.007 0.069 0.222 0.609 0.091 0.0011 150.7 158.2 0.184 1,519

(0.044) (0.124) (0.208) (0.235) (0.103) (0.0065) (1,310.4) (1,826.3) (0.159)Non-Importers 0.030 0.190 0.293 0.437 0.050 0.0003 65.0 46.5 0.150 3,690

(0.102) (0.239) (0.227) (0.292) (0.075) (0.0034) (1,141.0) (613.4) (0.160)Non-Exporting Plants

Importers 0.018 0.072 0.333 0.511 0.066 0.0004 40.6 41.6 0.184 3,993(0.084) (0.186) (0.214) (0.232) (0.088) (0.0031) (260.8) (461.2) (0.148)

Non-Importers 0.065 0.291 0.309 0.302 0.033 0.0002 15.6 12.6 0.132 20,262(0.158) (0.309) (0.252) (0.291) (0.078) (0.0056) (383.0) (183.9) (0.163)

Foreign-Owned PlantsImporters 0.008 0.070 0.170 0.651 0.099 0.0015 176.4 360.0 0.196 303

(0.045) (0.111) (0.183) (0.238) (0.108) (0.0054) (744.1) (3,726.2) (0.177)Non-Importers 0.023 0.130 0.208 0.555 0.083 0.0007 59.9 111.5 0.178 376

(0.086) (0.185) (0.205) (0.294) (0.108) (0.0038) (337.6) (843.4) (0.158)Domestic Plants

Importers 0.037 0.170 0.199 0.513 0.080 0.0012 115.4 100.6 0.179 2,178(0.115) (0.235) (0.203) (0.311) (0.104) (0.0065) (1103.6) (768.8) (0.170)

Non-Importers 0.071 0.329 0.276 0.291 0.033 0.0003 26.7 19.3 0.129 19,896(0.163) (0.303) (0.240) (0.301) (0.080) (0.0058) (623.5) (301.2) (0.164)

Panel B: Production vs. Non-Production WorkersProduction Workers Non-Production Workers

Less than Jr. College Less than Jr. CollegePrimary Primary High High Grad. Primary Primary High High Grad.

All PlantsImporters 0.016 0.078 0.328 0.544 0.035 0.002 0.018 0.168 0.566 0.245

(0.077) (0.182) (0.234) (0.264) (0.071) (0.037) (0.092) (0.204) (0.236) (0.232)Non-Importers 0.061 0.290 0.324 0.309 0.017 0.017 0.085 0.193 0.534 0.172

(0.156) (0.314) (0.267) (0.315) (0.065) (0.107) (0.232) (0.288) (0.352) (0.257)Exporting Plants

Importers 0.008 0.081 0.240 0.627 0.044 0.003 0.024 0.104 0.529 0.340(0.046) (0.144) (0.225) (0.266) (0.084) (0.026) (0.086) (0.157) (0.254) (0.271)

Non-Importers 0.030 0.206 0.314 0.429 0.021 0.013 0.053 0.133 0.543 0.258(0.104) (0.256) (0.247) (0.320) (0.060) (0.091) (0.166) (0.223) (0.322) (0.292)

Non-Exporting PlantsImporters 0.018 0.077 0.361 0.513 0.031 0.002 0.016 0.194 0.581 0.207

(0.086) (0.195) (0.229) (0.256) (0.066) (0.040) (0.094) (0.215) (0.227) (0.203)Non-Importers 0.067 0.305 0.326 0.287 0.016 0.018 0.091 0.206 0.532 0.154

(0.163) (0.322) (0.270) (0.309) (0.066) (0.110) (0.244) (0.298) (0.358) (0.245)Foreign-Owned Plants

Importers 0.009 0.084 0.181 0.686 0.041 0.004 0.023 0.074 0.498 0.401(0.044) (0.134) (0.200) (0.269) (0.085) (0.039) (0.075) (0.136) (0.276) (0.299)

Non-Importers 0.023 0.145 0.229 0.564 0.038 0.012 0.034 0.102 0.506 0.346(0.085) (0.208) (0.227) (0.336) (0.100) (0.081) (0.116) (0.198) (0.334) (0.330)

Domestic PlantsImporters 0.038 0.185 0.216 0.521 0.039 0.006 0.048 0.099 0.539 0.308

(0.118) (0.248) (0.225) (0.343) (0.086) (0.060) (0.146) (0.194) (0.296) (0.281)Non-Importers 0.073 0.346 0.290 0.275 0.016 0.021 0.107 0.168 0.529 0.175

(0.168) (0.315) (0.260) (0.320) (0.068) (0.120) (0.257) (0.287) (0.376) (0.273)

Notes: Standard deviations are in parentheses. The first column indicates current import status, where “importers”

denotes plants that import and “non-importers” captures plants that do not import in the current year. The first panel

pools all plants in all years. The second and third panel split the sample by export status, while the fourth and fifth

panels split the sample by the country of ownership. Specifically, foreign-owned plants are defined as those plants where at

least 10% of equity is held by foreign investors while domestic plants are defined as plants for which more than 90% of

equity is held by domestic investors.

6

greater percentage of skilled workers, invest more heavily in R&D and worker training,

and tend to use a greater percentage of non-production workers. Importantly, in any

subgroup we begin to see large differences in the hiring patterns of importing plants at

the high school level.

Panel B of Table 1 documents the percentage of total production or non-production

employment in each educational category. Again, importing plants are found to system-

atically hire more workers with high school degrees or better and fewer workers with

less education than high school within each of the two occupation categories. Although

importers always appear to be more skill-intensive on average within each occupation

category, the mechanism that drives the correlation between importing and skill-intensity

may differ between production and non-production workers. While the use of imported

materials might induce the adoption of new production processes which in turn requires

hiring more skilled production workers, importing might require substantial increases in

the number of non-production workers for trade related activities such as dealing with

customs agents or arranging shipments from foreign countries. Given the potential for

differences in the impact of importing on the demand for skilled workers across occupation

categories, we attempt to disentangle the causal impact of importing on the demand for

skilled workers within each occupation.

While Table 1 demonstrates the rich detail with which our data allow us to charac-

terize the distribution of skill, they also point to a potentially important limitation. In

particular, many Indonesian plants do not hire any workers with college or post-graduate

training in both occupation categories. As a result, defining a skilled worker as a “college

graduate” in this context would lead to eliminating a significant number of plants which

are wholly composed of unskilled workers. For this reason, we choose to define a skilled

worker as one with at least a high school degree. This definition not only allows us to bet-

ter exploit the observed variation in our data, but is also the natural metric for skill given

the context in which we conduct our empirical study. Indonesian high school graduates

are exposed to greater logical and quantitative training (Hendayana et al, 2008). Fur-

ther, the additional schooling allows high school students to refine their language skills in

standard Indonesian (Bahasa Indonesia) and, for many, the opportunity to learn to com-

municate effectively in English (Kam, 2006). These communication skills are particularly

important in this environment since Indonesian dialects vary widely across the country.

7

2.3 Instruments

We expect that the decision to import for any given plant is likely to be endogenously

determined with its decision to hire skilled labour. The identification strategy we outline

below relies on the presence of instruments. We consider a variety of instrument set com-

prised of location-specific transport costs, industry-specific measures of imported input

weight, industry-specific measures of the fraction of imports shipped by air, changes in

product-specific input and output tariffs and their interactions. We discuss the construc-

tion of each instrument in turn.

Since we do not observe transport costs to the port directly, we construct the measure

of transport cost for each plant as follows. To incorporate geographical information, we

first divide Indonesia into cells of one kilometer squared and assign a value of 1-10 to each

cell, where “10” is the highest cost (Steepness of Slope, Sea vs. Land). Then, we use

ArcGIS to find the least accumulative-cost path between any plant and its nearest port.

Finally, our measure of transport cost is obtained from the least accumulative-cost after

dividing it by the sample standard deviation.

For the second and third instruments, we rely on research that suggests that differences

in trade responses across industries can arise from differences in the nature of delivery

(e.g. air vs. water) or heaviness of the output. We extend this literature by combining

measures of the weight of imports to Indonesia or the fraction of imports shipped by

air, industry-by-industry, with an Indonesian import input-output table to construct an

industry-level measure of relative import weight.6 The intuition for the import weight

instrument comes directly from Cosar and Demir (2015) where we recognize that heavier

imports will be more costly to ship, particularly over land. Likewise, as Hummels and

Schaur (2013) argue, exporters pay a premium to ship goods by air for faster delivery.

For each industry j, the variables weightj and airsharej measure the weight of In-

donesian imports and the fraction of Indonesian imports shipped by air,

weightj = ln

(ocean weightjocean valuej

), and airsharej =

air valuejair valuej + ocean valuej

,

where air valuej denotes the value of air shipments to Indonesia in 2006 and ocean valuej

(ocean weightj) denotes the value (weight) of shipments to Indonesia by ocean in the

same year.7 We then combine this information with an import input-output table which

6The import input-output table is produced by BPS Indonesia.7Specifically, we compute these measures for shipments to Indonesia from the US and Europe. We

then take a simple average across both import sources. We thank Kerem Cosar who provided us with his

8

Figure 1: Import Weight and Airshare Instruments Across Industries0

12

3Im

po

rte

d I

np

ut

We

igh

t

15111 17220 24119 26900 31149 32210 35293 37205

Note: Each bar represents a distinct (5−digit) manufacturing industry code.

(a) Weight of Import Shipments

0.2

.4.6

Imp

ort

ed

In

pu

t A

irsh

are

15111 17220 24119 26900 31149 32210 35293 37205

Note: Each bar represents a distinct (5−digit) manufacturing industry code.

(b) Fraction of Import Shipments by Air

provides us with the share of imports purchased from each industry in Indonesia.8 Letting

shareij represent industry i’s import expenditure share on from industry j, our measures

of import weight and import air share in industry i are constructed as

Import weighti =∑j

shareij × weightj, Import airsharei =∑j

shareij × airsharej,

where∑

j shareij = 1 ∀ i by construction.

Figure 1 documents the variation in the fraction imports shipped by air and differences

in the weight of imported inputs across industries. It is clear that there exist substantial

differences across industries and, not surprisingly, industries which tend to import lighter

inputs are also more likely to have them shipped by air, where the correlation coefficient

between these instruments is -0.4.

For the fourth and the fifth instruments, we match each plant in our manufacturing

survey to product-level (5-digit ISIC) output and input tariffs constructed by Amiti and

Konings (2007). Full details of the construction of all instruments can be found in Ap-

pendix A. We use the change in output and input tariff rates between 1996 and 2001 as

our instrument. During our sample period Indonesia was broadly reducing tariffs across

manufacturing industries.

Naturally, we are concerned that the empirical estimates we find may be biased if the

Stata do files that compute variables weightj and airsharej from EU and US trade data sets.8BPS Indonesia produces detailed input-output tables measuring of total purchases (all sources) at

the industry-level or total domestic purchases. We construct measures import flows and import sharesby subtracting the information in the domestic input-output table from the comparable information inthe total input-output table. All input-output data is measured in 1995.

9

instruments we use are not exogenous. For the transport cost variable, it is possible that

plants with a high-return from importing will choose to locate closer to ports. To address

the potential concern for endogenous location choice, we also consider a sample of plants

which initially did not import in 1996. In this fashion, we can consider the impact of

transport costs (and tariffs) on plants who made their location decision well before they

began using imported materials. The transport cost variable along with import weight

and fraction of imports shipped by air comprise our ‘small instrument set.’ We treat these

instruments, particularly in conjunction of sample of initial non-importers, as our most

plausibly exogenous group of instruments.

We also consider an augmented instrument set which includes input and output tariff

changes along with the small instrument set and a full set of interactions between all of the

instruments. We initially exclude tariff changes from the small instrument set because it

is possible that tariff rate reductions are set to take advantage of industries where greater

importing will have larger impact on technological adoption and the relative demand for

skilled workers.9 In our IV regressions, we consider both specifications with and without

tariffs to investigate whether the tariff rates potentially lead to biased estimates.

3 A Simple Model of Importing, Selection and SBTC

Consider a simple two-country model of importing where home (Indonesian) firms consider

whether or not to import from abroad. Consumers have CES preferences and the market

structure is monopolistic competition. A home firm producing variety ω faces home

demand q(ω) = B(Zd)p(ω)−η where q is quantity demanded, p is the output price, η is

the elasticity of substitution, and Zd is a vector of observed variables that serve as a

demand shifter.10

Firms hire capital, skilled labor, and unskilled labor on competitive factor markets

and combine them with intermediate materials - purchased domestically or imported - to

produce output according to the production function

f(K,M,Ls, Lu, A, ϕ) = ϕKαkMαm{[ALs](σ−1)/σ + L(σ−1)/σu }αlσ/(σ−1), (1)

9As argued by Amiti and Konings (2007), however, the reductions are largely driven by the fact thatIndonesia is in the process of joining the WTO and, thus, are likely to be exogenous to the broaderpolitical or economic environment.

10We abstract from the possibility of exporting for the moment, though it will be straightforwardto extend our framework to allow for it. It is also straightforward to extend the model to allow forunobserved demand/cost shifters.

10

where K is capital, M is total intermediate materials, Ls is the number of skilled workers,

Lu is the number of unskilled workers, σ > 1 is elasticity of substitution between skilled

and unskilled workers, ϕ is a Hicks neutral productivity term, and A is a skilled labor

augmenting technology term. To consider the potential impact of importing on the relative

demand for skilled workers, we allow foreign imported inputs to affect the level of skilled

labor augmenting technology as

lnA(X,D, β) = Dβ +Xγ′ + U , (2)

where D is a dummy variable for the use of imported inputs, β is a firm-specific parameter

that captures the effect of importing on skill-biased technology A, X is a vector of ob-

servables (to be specified later), and U is a skill-biased technology shock. The skill biased

technology term A depends on X and the import decision D as in (2) but we assume that

importing does not affect the Hicks neutral technology level ϕ.11

For simplicity, we assume constant returns to scale technology where the plant’s

marginal cost is determined by

c(A,ϕ) = minLs,Lu,K,M

WsLs+WuLu+WkK+WmM subject to f(Ls, Lu, K,M,A, ϕ) ≥ 1 (3)

and (Ws, Wu, Wk, Wm) is a vector of factor prices. If the plant chooses to import, it incurs

a fixed import cost fm(Zc) where Zc contains a vector of observed variables that determine

the fixed cost. Then the plant’s net profit function is π(A,Zd, Zc, ϕ,D) = r(A,Zd, ϕ) −Dfm(Zc), where r(A,Zd, ϕ) = max

qpq − c(A,ϕ)q. A firm will import whenever the net

profit from importing is greater than the net profit achieved using domestic materials

alone,

∆π(X,Zd, Zc, ϕ, β) := π(A(X, 1, β), Zd, Zc, ϕ, 1)− π(A(X, 0, β), Zd, Zc, ϕ, 0) ≥ 0. (4)

Note that ∆π(X,Zd, Zc, ϕ, β) is strictly increasing in β and ϕ.

Suppose we allow β to vary across plants as

β = ¯β + ε,

where ¯β is the mean of β and ε is the plant-specific return to importing. Then, for each

11It is also possible to allow importing to improve Hicks-Neutral productivity. We abstract from thispossibility only for expositional simplicity.

11

Figure 2: Importing and Skill Demand

β∗

ϕ

(a) Marginal Importer’s β∗

S1 − S0

ϕ

(b) Marg. Importer’s S1−S0

E[S1 − S0|X = x, UD = p]

p

(c) MTE

value of X, Zd, Zc, and ϕ, it is straightforward to determine a threshold value of β that

induces firms to start importing by the condition ∆π(X,Zd, Zc, ϕ, β∗) = 0. This threshold

value β∗ depends on X, Zd, Zc, and ϕ. Naturally, firms with low initial productivity ϕ

will require larger values of β to justify importing.12

Under the assumption of heterogeneous returns to importing, we can illustrate the

selection mechanism by considering the locus of β’s for the marginal importer. Figure

1(a) demonstrates that this locus is strictly decreasing in initial Hicks-neutral productivity,

ϕ, while fixing the value of X, Zd, and Zc. Firms with low Hicks-neutral productivity

(i.e., low value of ϕ) choose to import only if they receive high idiosyncratic returns from

importing (i.e., high value of β).

3.1 Selection, SBTC and Skill Demand

Consider the first order conditions from cost minimization problem (3). Denote the log

of the demand for skilled workers relative to unskilled workers by SD = ln(Ls/Lu) for

D ∈ {0, 1}, where the subscript D indicates its dependence on import status. Given

market wages, the relative demand for skilled workers is determined by equating the ratio

of the marginal product of skilled and unskilled workers to the ratio of their wages as

SD = (σ − 1) lnA− σ ln(Ws/Wu)

= Dβ +Xγ′ + U

= D(β + ε) +Xγ′ + U, (5)

12This argument is analogous to that in Lileeva and Trefler (2010) which studies the heterogeneousreturn to exporting on Hicks-neutral productivity.

12

where (β, β, ε, γ, U) are ( βσ−1

,¯βσ−1

, εσ−1

, γσ−1

, Uσ−1

) and X subsumes ln(Ws/Wu). In

our context, importing is an endogenous decision because the import decision D and

skill-biased technology shock U are likely to be correlated. Moreover, as indicated in

(4), plants with a greater ability to adopt skilled-biased technology (i.e., high value of

β) will self-select into importing because they will achieve larger productivity gains from

importing. Therefore, β and D are also correlated.

Because of this positive sorting on the gain from importing, we would expect that the

change in skill demand will be greater among plants that choose to import relative to

non-importers should they have started importing. As illustrated in Figure 1(b), similar

to Figure 1(a), the effect of importing on demand for skill for the marginal importer is a

decreasing function of Hicks-neutral productivity ϕ.

3.2 The Marginal Treatment Effect

We are interested in identifying the impact of importing on the demand for skilled labor,

β, which may vary across plants. Imbens and Angrist (1994) show that, under certain

conditions, using a single dummy instrument, an IV estimator identifies the local average

treatment effect (LATE), or the average value of β among plants who are induced to

change their import choices by the instrument. When multiple dummy instruments are

used, an IV estimator identifies a weighted average of the instrument-specific LATEs.

Therefore, an IV estimator provides an estimate of an interpretable quantity even when

the effect of importing on the demand for skilled workers is heterogenous across plants,

although the LATE is generally different from the average value of β.

To evaluate the heterogeneous impact of importing on the demand for skill, we use

the framework developed by Heckman and Vytlacil (1999, 2005, 2007a, 2007b) as follows.

The relative demand for skilled labor SD for D = 0 or 1 can be written as

S1 = µ1(X) + U1 and S0 = µ0(X) + U0, (6)

respectively, where, allowing for the average value of β to depend on X in (5), µ1(X) ≡E[S1|X] = β(X) + Xγ′ and µ0(X) ≡ E[S1|X] = Xγ′ while U1 = ε + U and U0 = U .

The impact of importing on the demand for skilled workers depends on the plant-specific

ability to adopt foreign technology embedded in imports since S1−S0 = β(X) +U1−U0.

To derive an empirical specification for the decision to import, let Z ≡ (Zd, Zc, ϕ)

and write (4) as ∆π(Z, β) ≡ ∆π(X,Zd, Zc, ϕ, β), where we assume that Z contains all of

13

the components of X for notational brevity. The Z vector also includes the instrumental

variables that are excluded from X. Define the latent variable, D∗, as D∗ = ∆π(Z, β) =

µD(Z) − V , where µD(Z) = E[∆π(Z, β)|Z] is a deterministic function of observable

variables Z while V = ∆π(Z, β)−µD(Z) is a mean-zero unobserved stochastic component.

Then, we have a latent variable model of importing:

D∗ = µD(Z)− V, D = 1 if D∗ ≥ 0, D = 0 otherwise. (7)

A plant imports if D∗ ≥ 0; it does not import otherwise. We assume that the distribu-

tion of V , denoted by FV , is continuous and strictly increasing and, furthermore, that

(U0, U1, V ) is statistically independent of Z given X.13

Let P (Z) denote the probability of importing conditional on Z so that P (Z) ≡Prob(µD(Z) > V ) = FV (µD(Z)). P (Z) is called the propensity score. Define UD ≡FV (V ), and the random variable UD is uniformly distributed on [0, 1] by construction.

Because V is the unobserved component of the net benefit of importing, UD represents

the quantiles of the unobserved net benefit from importing. Then the import decision (7)

is alternatively written as D = 1 if P (Z) ≥ UD and D = 0 otherwise.

We define the marginal treatment effect (MTE) as

∆MTE(x, p) = E[S1 − S0|X = x, UD = p] = β(x) + E[U1 − U0|X = x, UD = p]. (8)

This is the average effect of importing on skilled demand for plants with X = x and

UD = p. Because a plant is indifferent between importing and not importing when

P (Z) = UD, the MTE captures the mean impact from importing on the demand for

skilled labor among plants with X = x and P (Z) = p when the realization of UD is such

that the plant is just indifferent between importing and not importing.

Estimating the MTE for each value of UD = p within the support of P (Z), we are able

to construct the empirical counterpart to the locus of returns as outlined in Figure 1(c).

Compared to Figures 1(a) or (b), the x-axis is measured in import probabilities rather than

productivity in Figure 1(c) since we allow firms to differ in many dimensions rather than

only productivity. Propensity scores are a natural metric to summarize those observed

differences in a single dimension. When firms self-select into importing based on their

unobserved benefits from importing, we expect the MTE curve to be downward sloping

13The latter is implied by the independence and monotonicity assumptions of Imbens and Angrist(1994) as shown by Vytlacil (2002).

14

because, if firms choose to import even if their observed characteristics suggest that they

were not likely to import (i.e., the low value of P (Z) = p), the unobserved component of

net benefit from importing must be high (i.e., the high value of E[U1−U0|X = x, UD = p]

in (8)). In contrast, we expect the MTE curve to be flat in the absence of self-selection

based on the unobserved benefit from importing.

The MTE further allows us to compute all the conventional treatment parameters, such

as the ATE, the TT, and the TUT, as weighted averages of the MTE, each computed

with a different weighting function (see Heckman and Vytlacil (2005, 2007a, 2007b)).

We estimate the MTE and treatment parameters following a procedure similar to that

of Carneiro, Heckman, and Vytlacil (2011). Because the support of P (Z) for each value

of X is small, as in Carneiro, Heckman, and Vytlacil (2011), we assume that (X,Z) is

independent of (U1, U0, UD). Then, the MTE can be identified within the support of

P (Z) as ∆MTE(x, p) = β(x) + E[U1 − U0|UD = p], where the term β(x) represents the

average treatment effect when X = x while E[U1−U0|UD = p] represents the component

of the MTE that depends on UD. Furthermore, because X is a high-dimensional vector,

allowing the value of β to depend on all variables in X leads to imprecise estimates

of β(X). We set β(X) = X ′θ, where X contains the log of each skill ratio in 1996,

log(Ljs/Lju)1996, and includes dummies for plants that did not hire any skilled or unskilled

workers, djs = 1(Ljs = 0) and dju = 1(Lju = 0) in 1996.14 Then,

E[S|X = x, P (Z) = p] = x′γ + px′δ +K(p), ∆MTE(x, p) = x′δ +K ′(p), (9)

where K(p) = E[U1−U0|UD ≤ p]p and K ′(p) is the first derivative of K(p). We estimate

γ, δ, and K(p) by a partially linear regression of S on X and P (Z) (Robinson, 1988) with

local polynomial regressions as described in Appendix B.

4 Empirical Results

4.1 Definitions of Variables and Sample Selection

All outcome variables and most explanatory variables are measured in 2006. The lagged

value of outcome variables are also included in the set of explanatory variables so that

14In our preliminary investigation, when we estimated (9) by setting X equal to all variables in Xexcept for the local wage ratios, industry dummies, and province dummies, we found that the interactionterms with other variables in X were rarely significant across different specifications.

15

Table 2: Definitions of the Variables

Var. Definition

S The log of the ratio of skilled workers to unskilled workers in 2006 in occupation j ∈ {p, n}, ln(Ljs/L

ju

)06

,

the share of skilled workers in the total wage bill in 2006 in occupation j ∈ {p, n},(WsL

js/(WsL

js +WuL

ju))06

.

D Equal to one if plant imports materials from abroad in 2006; zero otherwise.X Export dummy, capital stock, a measure of Hicks-neutral productivity, foreign ownership dummy, a dummy

for positive R&D expenditures, a dummy for positive training expenditures, the log of the ratio of skilledworkers’ wages to unskilled workers’ wages in 1996 and in 2006 in the region where a plant locates, the1996 value of the outcome variables denoted by ln

(Ljs/L

ju

)96

and(WsL

js/(WsL

js +WuL

ju))96

for j ∈ {p, n},a dummy for no hiring of skilled workers or unskilled workers in occupation j ∈ {p, n} in 1996 denoted by

djs,96 := 1(Ljs = 0) or dju,96 := 1(Lj

u = 0), respectively, TFP measure constructed by the Levinsohn andPetrin method, 3-digit ISIC industry dummies, and province dummies.

Z\X Transport cost to the nearest port, the average weight of Indonesian imports per value in industry j,the fraction of Indonesian imports shipped by air in industry j, a change in output and input tariffrates at 5-digit ISIC level between 1996 and 2001.

Notes: A skilled worker is defined as a worker with high school eduction and an unskilled worker isdefined as a worker without high school education. Occupation categories “p” and “n” denoteproduction workers and non-production workers, respectively. All variables are measured in 2006 unlessstated otherwise.

our sample consists of plants that are present in both the 1996 and 2006 data sets. The

definitions of variables and their descriptive statistics are reported in Tables 2 and 3.

We consider four different outcome variables for SD. Our first two measures, ln (Lps/Lpu)06

and ln (Lns/Lnu)06, directly capture the number of skilled workers within each occupation

category where Ljs and Lju are the number of skilled workers and unskilled workers, re-

spectively, employed in occupation j ∈ {p, n}, the superscripts “p” and “n” distinguish

production and non-production workers, respectively, and the subscript “06” indicates

that a variable is measured in 2006. As before, we define a skilled worker as one with at

least a high school degree. The next two outcome variables, (WsLps/(WsL

ps +WuL

pu))06

and (WsLns/(WsL

ns +WuL

nu))06, measure the ratio of skilled workers’ wages to the total

wage bill in each occupation category, where Ws and Wu represent local market wages for

skilled workers and unskilled workers, respectively. A non-trivial number of plants do not

hire any skilled workers in each occupation. When the log of the ratio of skilled workers to

unskilled workers is used as an outcome variable, we simply ignore these plants for which

we cannot compute the outcome variable. However, this omission itself may generate

selection bias. Therefore, we consider the skilled worker’s wage share in the total wage

bill for occupation j, WsLjs/(WsL

js +WuL

ju), as an alternative outcome variable, which is

related to changes in the skill-biased technology parameter, A/(1 +A), when we take the

16

Table 3: Descriptive Statistics

Full Sample Sample of Initial Non-importersD = 0 D = 1 D = 0 D = 1

Explanatory Variable(a) Mean S.D. Mean S.D. Mean S.D. Mean S.D.TC 0.812 0.858 0.529 0.671 0.804 0.818 0.632 0.807Air 0.093 0.074 0.122 0.093 0.088 0.066 0.117 0.090Wgt 0.884 0.687 0.891 0.629 0.896 0.694 0.880 0.654Export 0.183 0.387 0.497 0.500 0.172 0.378 0.436 0.497Capital 13.734 1.904 15.468 2.204 13.546 1.773 15.126 2.019Hicks-neutral ϕ 5.259 0.598 5.638 0.638 5.221 0.568 5.550 0.637Foreign 0.020 0.138 0.106 0.308 0.016 0.125 0.089 0.285R&D 0.065 0.247 0.205 0.404 0.057 0.232 0.178 0.383Training 0.330 0.470 0.640 0.480 0.301 0.459 0.591 0.493log(Ws/Wu)06 0.444 0.176 0.448 0.155 0.440 0.177 0.451 0.154log(Ws/Wu)96 0.365 0.134 0.349 0.119 0.366 0.134 0.350 0.126ln(Lps/L

pu)96 -0.676 1.374 -0.253 1.549 -0.742 1.354 -0.373 1.493

dpu,96 0.019 0.136 0.053 0.224 0.016 0.126 0.056 0.231

dps,96 0.314 0.464 0.089 0.284 0.342 0.475 0.149 0.356

No. Obs. 5716 4187

Full Sample Sample of Initial Non-importersD = 0 D = 1 D = 0 D = 1

Outcome Variable(b) Mean S.D. Mean S.D. Mean S.D. Mean S.D.ln(Lps/L

pu)06 -0.393 1.684 0.578 1.673 -0.520 1.661 0.454 1.653

ln(Lns /Lnu)06 0.671 1.319 1.268 1.375 0.588 1.317 1.258 1.500

(WsLps/(WuL

pu +WsL

ps))06 0.427 0.353 0.659 0.311 0.400 0.348 0.608 0.334

(WsLns p/(WuLnu +WsLns ))06 0.785 0.332 0.862 0.241 0.779 0.338 0.825 0.295

Notes: (a) The sample statistics for the explanatory variables that are used to estimate the decisions to import in Table

E.2. (b) The sample statistics for the outcome variables that are used to estimate the skill demand equation (9).

limit of σ → 1 in production function (1), i.e., Cobb-Douglas case.

To control for the initial level of skilled biased technology in 1996, X includes the

lagged value of the outcome variable in 1996, denoted by using the subscript “96” in place

of “06,” dummy variables for plants that did not hire any skilled or unskilled workers

in each occupation in 1996, denoted by djs,96 and dju,96 for j = p, n, and the relative

wage ratios in 1996, denoted by ln(Ws/Wu)96.15 We also include capital stock in X to

control for differences across size or capital-intensity. In addition, X contains the plant’s

current export status, our estimate of Hicks-neutral productivity ϕ, the local skilled-

unskilled wage ratio,16 a large set of dummy variables to capture differences across foreign

ownership, R&D expenditures, worker training expenditures, industries and provinces.

Using production function (1), we estimate a model-consistent measure of Hicks-neutral

productivity ϕ based on the frameworks developed by Olley and Pakes (1996), Levinsohn

15Here, we assume that the lagged outcome variable, say, ln (Lps/L

pu)96 takes a value equal to zero when

either Lps = 0 or Lp

u = 0.16The wage ratio is measured through a series of Mincer regressions described in Appendix A. We

proceed in this fashion so to isolate the local difference in wages due to education alone, rather than havedifferences in the wage ratio reflect differences in demographics, experience, etc across regions.

17

and Petrin (2003), Ackerberg, Caves and Frazer (2015) and Gandhi, Navarro and Rivers

(2013) as described in Appendix C. For robustness, we also estimated a conventional

measure of TFP from a standard Cobb-Douglas production function and used this in

place of our Hicks-neutral productivity measure. Finally, Z\X includes our instruments.

4.2 IV Results

4.2.1 Production Workers and Non-Production Workers

Table 4 presents the results from estimating equation (5) by OLS and IV when we use

the log of the ratio of skilled production to unskilled production workers as its dependent

variable. Columns (1)-(4) report the results for the full sample while columns (5)-(8)

report those for the sample of initial non-importers. Using only the sample of plants that

were not importing in 1996 allows us to consider the impact of transport costs on plants

who made their location decision before they began using imported materials.

As reported in columns (1) and (5), the OLS point estimates suggest that importing

significantly increases the relative demand for skilled workers within the production oc-

cupation by 28-32 log basis points. Columns (2)-(4) and (6)-(8) report the results from

IV regressions using different sets of IVs. For columns (2) and (6), we use a small set of

instrumental variables: internal transport costs, import heaviness and import air share,

as defined in Section 2.3. Columns (3) and (7) use an expanded instrument that includes

the all variables in the ‘small instrument set’ but also incorporates the change in input

and output tariffs along with interactions between all instruments. Last, columns (4) and

(8) again use the full instrument set, but instead consider the import share, defined as the

percentage of materials which are imported, as the treatment variable rather than import

status.17 First stage regressions indicate that transportation costs and import air share

are always a strong predictors of import behavior. See Table E.1 in Appendix E.

Across different sets of IVs and different samples, our IV estimates are larger than

the OLS estimate by an order of magnitude; for the full sample, the point estimates on

import status range between 3.33-4.81 indicating that importing increases the relative

demand for skilled workers within the production occupation by at least 333 log basis

points. In a “standard” case where we assume that there is no heterogeneity in β in (5),

the finding that the IV estimate is much larger than the OLS estimate could be viewed

17We have considered all other specifications using the import share treatment variable. Since theresults were very similar to those reported in columns (4) and (8) of Tables 2, 3 and 4, we chose to omitthem from the manuscript. They are available upon request.

18

Table 4: Skill Demand Equation for Production Workers

Dependent Variable: (Lps/Lpu)06

Full Sample Sample of Initial Non-importers (D96 = 0)(1) (2) (3) (4) (5) (6) (7) (8)

OLS IV using IV using IV using OLS IV using IV using IV usingsmall IV set full IV set full IV set small IV set full IV set full IV set

Import Status 0.279*** 3.331*** 4.809*** 0.321*** 3.435*** 2.153**[0.059] [0.682] [1.256] [0.102] [1.246] [0.928]

Import Share 5.483*** 1.039[1.858] [1.266]

Export Status 0.014 -0.469*** -0.522*** -0.300** -0.060 -0.265** -0.190** -0.073[0.057] [0.131] [0.166] [0.128] [0.072] [0.116] [0.097] [0.080]

Capital 0.123*** 0.039 0.058** 0.106*** 0.136*** 0.069** 0.092*** 0.127***[0.013] [0.025] [0.026] [0.020] [0.017] [0.030] [0.024] [0.020]

Hicks-neutral, ϕ -0.251*** -0.339*** -0.334*** -0.383*** -0.260*** -0.307*** -0.279*** -0.266***[0.040] [0.058] [0.067] [0.076] [0.051] [0.068] [0.060] [0.056]

Foreign-Owned 0.070 -0.528*** -0.576** -0.217 0.088 -0.214 -0.085 0.201[0.101] [0.201] [0.253] [0.212] [0.150] [0.240] [0.194] [0.157]

R&D 0.043 -0.176 -0.199 -0.050 0.015 -0.172 -0.095 0.002[0.072] [0.113] [0.139] [0.108] [0.100] [0.146] [0.124] [0.107]

Training 0.234*** 0.087 0.074 0.157** 0.220*** 0.138* 0.181*** 0.243***[0.050] [0.073] [0.086] [0.071] [0.059] [0.080] [0.070] [0.062]

ln (Ws/Wu)06 -0.024 -0.169 -0.237 -0.002 -0.077 -0.243 -0.202 -0.113[0.112] [0.189] [0.216] [0.194] [0.133] [0.205] [0.188] [0.177]

ln (Ws/Wu)96 -0.489*** -0.448** -0.555** -0.535** -0.681*** -0.790*** -0.750*** -0.737***[0.174] [0.228] [0.261] [0.223] [0.206] [0.241] [0.222] [0.213]

ln(Lps/Lpu)96 0.387*** 0.342*** 0.372*** 0.368*** 0.357*** 0.337*** 0.344*** 0.360***

[0.018] [0.026] [0.027] [0.023] [0.022] [0.028] [0.025] [0.023]dpu,96 0.523*** 0.310 0.210 0.260 0.188 -0.038 0.050 0.173

[0.141] [0.207] [0.262] [0.238] [0.198] [0.265] [0.230] [0.204]dps,96 -1.033*** -0.858*** -0.964*** -0.982*** -1.025*** -1.000*** -1.000*** -0.996***

[0.064] [0.086] [0.089] [0.078] [0.073] [0.083] [0.078] [0.075]Import Status1996 -1.697***

[0.511]Import Share1996 -2.197**

[0.871]No. Obs 4,631 4,288 4,285 4,095 3,297 3,061 3,059 3,006R-squared 0.385 — — — 0.337 — — —

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs.

19

Table 5: Skill Demand Equation for Production Workers

Dependent Variable:(WsL

ps/(WuL

pu +WsL

ps))06

Full Sample Sample of Initial Non-importers (D96 = 0)(1) (2) (3) (4) (5) (6) (7) (8)

OLS IV using IV using IV using OLS IV using IV using IV usingsmall IV set full IV set full IV set small IV set full IV set full IV set

Import Status 0.048*** 0.661*** 0.929*** 0.037** 0.761*** 0.524***[0.010] [0.125] [0.222] [0.016] [0.256] [0.200]

Import Share 1.292*** 0.536*[0.366] [0.316]

Export Status 0.018* -0.086*** -0.096*** -0.063** 0.012 -0.044* -0.026 -0.001[0.010] [0.025] [0.032] [0.027] [0.013] [0.025] [0.021] [0.015]

Capital 0.024*** 0.008* 0.011** 0.018*** 0.026*** 0.013*** 0.017*** 0.023***[0.002] [0.004] [0.004] [0.004] [0.003] [0.005] [0.004] [0.004]

Hicks-neutral, ϕ -0.010 -0.022** -0.019* -0.034*** -0.012 -0.023** -0.018* -0.018*[0.006] [0.009] [0.010] [0.012] [0.008] [0.011] [0.010] [0.009]

Foreign-Owned -0.006 -0.115*** -0.125*** -0.097** -0.001 -0.082* -0.054 -0.008[0.017] [0.034] [0.043] [0.044] [0.029] [0.050] [0.041] [0.035]

R&D 0.020* -0.020 -0.022 0.001 0.034** -0.000 0.012 0.023[0.012] [0.020] [0.024] [0.021] [0.017] [0.026] [0.022] [0.019]

Training 0.045*** 0.020 0.020 0.027** 0.049*** 0.033** 0.040*** 0.050***[0.009] [0.013] [0.014] [0.013] [0.010] [0.015] [0.013] [0.012](

WsLps

(WuLpu+WsL

ps)

)96

-0.403*** -0.357*** -0.390*** -0.405*** -0.404*** -0.378*** -0.387*** -0.408***

[0.017] [0.026] [0.027] [0.025] [0.021] [0.030] [0.026] [0.023]dpu,96 -0.055*** -0.079** -0.116*** -0.127*** -0.057* -0.087* -0.074* -0.067**

[0.019] [0.032] [0.042] [0.042] [0.030] [0.046] [0.039] [0.034]dps,96 -0.430*** -0.383*** -0.418*** -0.435*** -0.429*** -0.418*** -0.421*** -0.431***

[0.016] [0.023] [0.023] [0.022] [0.019] [0.024] [0.022] [0.021]Import Status1996 -0.321***

[0.087]Import Share1996 -0.514***

[0.166]No. Obs 6,258 5,827 5,824 5,562 4,626 4,328 4,326 4,254R-squared 0.453 — — — 0.415 — — —

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs.

20

as puzzling since the OLS bias may likely be upward in this case. When the coefficient

β is random, however, the finding of the large IV estimate is less puzzling because the

IV estimator identifies the local average treatment effect. Our results suggest that, on

average, only those plants with very high values of β—interpreted as plants with a better

ability to adopt skill biased technology—choose to change their import status. Moreover,

it is important to recall the context of this estimate. In our sample, as reported in Table

1, the ratio of skilled to unskilled workers among importers is more than double that of

the average non-importer. Columns (4) and (8) check the robustness of our result using

import share, a continuous measure of import intensity, in place of the import status

variable. In either case, we find an estimated coefficient which is very similar both in

terms of magnitude and significance.

The control variables in Table 4 generally report consistent and intuitive coefficients.

The estimated coefficients on plant-level export status are often negative, which reflects

the fact that Indonesia has a comparative advantage in unskilled-labor intensive goods.

The significant positive capital and training coefficients indicate both capital and training

are complementary to hiring skilled labor. On the other hand, foreign ownership is often

negatively associated with the demand for skilled labor, suggesting that foreign ownership

is a substitute for skill-intensive production processes (e.g. by offshoring the skill-intensive

portion of production abroad). The estimated coefficient on Hicks-neutral productivity

ϕ is negative, which suggests a trade-off between the adoption of skill-biased technology

and the adoption of technology that is unbiased across skill differences. The coefficient on

relative wages in 2006 is negative, as expected, but insignificant. The estimated coefficient

on the lagged value of the outcome variable is positive and significant, reflecting either the

persistence of unobserved characteristics that affect the plant’s demand for skilled labor

or the presence of adjustment costs associated with changing the plant’s skill ratio.

Table 5 reconsiders the results from estimating equation (5) using the skilled labor

share of the production wage bill, (WsLps/(WuL

pu +WsL

ps))06, as the dependent variable

and repeat each of our benchmark regressions. We again find that importing is predicted to

have a large, positive and significant impact on the relative demand for skilled production

workers in each column. Moreover, the IV estimates are substantially larger than the

OLS estimates, consistent with our previous findings.

Table 6 reports a number of robustness checks for the IV results for both dependent

variables. First, ignoring the endogeneity of the plant’s export decision may lead to the

bias in the estimated coefficient on import status given the evidence that importing and

exporting are closely related activities (see Kasahara and Lapham, 2013). To examine

21

Table 6: Robustness Checks. Skill Demand for Production Workers

(1) (2) (3) (4) (5) (6)Sample of Sample of IV Import IV Import Replace ϕ Replace ϕ

Export96 = 0, Export96 = 0, and Export and Export with TFP; with TFP;Export06 = 0; Export06 = 0; using full using full IV Import IV Import

IV Import IV Import Import IV Import IV using small using fullusing small using full set, full Export set and IV set IV set

IV set IV set full Export IV set andIV set Export96

dummy

Dependent Variable: ln(Lps/L

pu

)06

Import Status 3.423*** 3.031*** 1.169** 1.118** 3.032*** 2.671***[1.092] [0.956] [0.485] [0.509] [0.635] [0.570]

Export Status -1.282*** -2.044*** -0.465*** -0.410***[0.332] [0.543] [0.125] [0.114]

TFP 0.082* 0.075*[0.044] [0.042]

Export Status96 0.799***[0.260]

No. of Obs. 2,747 2,747 3,690 3,690 4,290 4,287

Dependent Variable:(WsL

ps/(WuL

pu +WsL

ps))06

(1) (2) (3) (4) (5) (6)Import Status 0.792*** 0.464*** 0.318*** 0.307*** 0.629*** 0.587***

[0.214] [0.161] [0.095] [0.098] [0.119] [0.110]Export Status -0.202*** -0.331*** -0.082*** -0.076***

[0.062] [0.104] [0.024] [0.022]TFP 0.023*** 0.022***

[0.008] [0.008]Export Status96 0.132***

[0.051]No. of Obs. 4,041 4,041 5,033 5,033 5,829 5,826

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs. The full sample is used in each column.

22

this issue, we estimate the skill equation using the subsample of plants that export neither

in 1996 nor 2006, as well as by instrumenting both import and export status using the

subsample of plants that did not export in 1996; we find that the point estimates for

importing remain significantly positive in all cases as reported in columns (1)-(4) of Table

6. Second, across different sets of IVs, we continue to find a significant and positive

effect of importing when we use conventional TFP in place of our estimated Hicks-neutral

productivity as reported in columns (5) and (6).18

Given our results for production workers, we next examine the effect of importing on

the relative skill demand among non-production workers. Specifically, we repeat the same

set of exercises as in Tables 4-6 using the variables defined in terms of non-production

workers in place of those defined in terms of production workers. Table 7 reports a

summary of the results. Across all columns of Table 7, importing appears to have a

large significant impact on demand for skilled non-production workers when we study the

sample which includes both 1996 importing and non-importing plants. However, when we

examine the sample of initial non-importers the impact of importing is never estimated to

be significantly different than zero. As demonstrated in the bottom panel of Table 7, this

pattern persists regardless of which dependent variable we use as the outcome measure.

4.2.2 Discussion

Although we have presented evidence that importing leads to an increase in the demand

for skill - clearly among production workers and less so among non-production workers

- the mechanism behind this result has been largely uninvestigated thus far. As we

discussed, one plausible mechanism is that importing induces the adoption of skill-biased

technology. While there is no direct data on foreign technology adoption, our data set

includes a variable which captures whether a plant adopts a standardized production

process, such as those recognized by the International Organization for Standardization

(ISO) or the International Electrotechnical Commission (IEC).19 The use of standards

18Hicks-neutral productivity is estimated to take a negative coefficient in Tables 4-5, while our naivelyestimated TFP takes the opposite sign in column (1) of Table 6. While these results may seem con-tradictory, they are in fact exactly what we should expect in this instance. By ignoring the skill-biasedcomponent of productivity, the conventional TFP measure confuses both the skill-biased and Hicks-neutral components and, as a result, is likely to be positively correlated with the demand for skilledlabour. In contrast, the Hicks-neutral productivity term we estimate disentangles these two componentsof productivity. Plants with larger values of skill-biased productivity will naturally be more likely to havesmaller measured Hicks-neutral productivity, ceteris paribus.

19Specifically, the survey question asks “Does this establishment use standard of production process?”with the following list of standards: ISO (International Organisation for Standardization), IEC (In-

23

Table 7: Skill Demand Equation for Non-Production Workers

Dependent Variable: ln (Lns /Lnu)06

(1) (2) (3) (4) (5) (6)OLS IV using IV using IV using IV using IV using

small IV set full IV set full IV set, full IV set full IV set,Import96 = 0 Import96 = 0

Import Status 0.126 2.205*** 1.970** 1.256*[0.121] [0.643] [0.804] [0.662]

Import Share 2.269* 1.786*[1.261] [0.968]

Export Status 0.089 -0.145 -0.043 -0.037 0.092 0.036[0.095] [0.129] [0.121] [0.109] [0.097] [0.098]

No. Obs 1,684 2,244 2,241 1,564 2,122 1,536

Dependent Variable: ln (Lns /Lnu)06

(7) (8) (9) (10) (11) (12)Sample of Sample of IV Import IV Import Replace ϕ Replace ϕ

Export96 = 0, Export96 = 0, and Export and Export with TFP; with TFP;Export06 = 0; Export06 = 0; using full using full IV Import IV Import

IV Import IV Import Import IV Import IV using small using fullusing small using full set and full set, full Export IV set IV set

IV set IV set Export IV IV set andIV set Export96

dummyImport Status 1.256 0.898 1.248** 1.273** 2.035*** 1.705***

[0.804] [0.680] [0.494] [0.506] [0.613] [0.549]Export Status -0.979*** -1.356** -0.147 -0.098

[0.362] [0.595] [0.126] [0.115]TFP -0.009 -0.017

[0.048] [0.046]No. of Obs. 1,409 1,409 1,899 1,899 2,244 2,241

Dependent Variable: (WsLns /(WuLnu +WsLns ))06(13) (14) (15) (16) (17) (18)OLS IV Import IV Import Sample of IV Import Replace ϕ

using small using full Export06 = 0, and Export96 = 0 with TFP;IV set IV set Export96 = 0; using full IV Import

IV Import Import IV using fullusing full set and IV set

IV set full ExportIV set

Import Status -0.010 0.321*** 0.269*** 0.178 0.130 0.272***[0.009] [0.111] [0.101] [0.169] [0.098] [0.097]

Export Status 0.019** -0.039* -0.030 -0.392*** -0.031[0.009] [0.022] [0.020] [0.126] [0.020]

TFP 0.005[0.007]

No. of Obs. 6,258 5,827 5,824 4,041 5,033 5,826

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs. The full sample is used in each column.

24

Table 8: Skill Demand, Import Decisions, and the Adoption of Standardized ProductionProcesses

Sample All Importers Sample of Initial Non-importersExport06 =Export96 = 0 Export06 =Export96 = 0

IV OLS IV standards IV standards OLS IV standards IV standardsusing small IV set using full IV set using small IV set using full IV set

Dependent Variable: ln(Lps/Lpu)06

(1) (2) (3) (4) (5) (6)Standards 0.033*** 1.181*** 0.869*** 0.043*** 0.911** 0.794***

[0.013] [0.407] [0.246] [0.015] [0.355] [0.244]

Dependent Variable: ln(Lns /Lnu)06

(7) (8) (9) (10) (11) (12)Standards 0.046 1.537 1.174* 0.017 1.213 1.149*

[0.084] [0.988] [0.706] [0.102] [0.928] [0.673]

Dependent Variable: Dummy Variable for Standardized Production in 2006(13) (14) (15) (16) (17) (18)

Import Status 0.074*** 0.520** 0.344* 0.102*** 1.201* 0.268[0.021] [0.208] [0.183] [0.035] [0.721] [0.406]

No. Obs. 4,303 4,041 4,041 3,462 3,263 3,263

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs. The full sample is used in each column.

may allow for improved coordination with foreign suppliers, facilitating the adoption of

foreign skill-biased technology.

In the top panel of Table 8, we estimate the effect of standards on the demand for

skilled production workers using the same specifications as in columns (1)-(3) and (5)-(7)

of Table 4, but replacing the import dummy with a dummy for standardization, where

we control for past export status since standardization is likely to be closely associated

with exporting activities.20 The results in columns (1)-(6) indicate that the adoption of

standards significantly increases the demand for skilled production workers. The middle

panel of Table 8 repeats this exercise for non-production workers, but, in contrast, we

find little statistically significant evidence to support the hypothesis that the adoption

of standards has any impact on the demand for skilled non-production workers. In this

ternational Electrotechnical Commission), ITU (International Telecommunication Union), CAC (CodexAlimentarius Commission), AFNOR (Association Francaise de Normalisation), ANSI (American NationalStandard Institute), BIS (Bureau of India Standard), BSI (British Standards Insitution), DIN (DeutshesInstitute for Nonnung ev), JISC (Japanese Industrial Standartds Commitee), SAL (Standards Australia),SNI (Standar Nasional Indonesia), ASTM (American Society for Testing and Material), ASME (AmericanStandard of Mechanical Engineering), and NFPA (National Fire Protection Association). Unfortunately,no further information on which standards are used is available.

20Transport costs are typically strong predictors of the use of standards. See Tables E.1- ?? in AppendixE.

25

sense, our evidence suggests that if importing leads to skill upgrading through SBTC, it

does so through production workers rather than non-production workers. Last, we also

estimate a linear probability model for standardization while instrumenting import (and

export) decisions using IVs in the bottom panel of Table 8 where we again include the

full set of control variables documented in Table 4. We find that importing significantly

increases the probability of the adoption of standards. Together our findings reflect that

importing is likely associated with the adoption of technology and that the adoption of

technology in turn leads to an increase in the demand for skilled production labor.

Although we argue that the above results present substantial evidence in favor of a

SBTC interpretation of our results, there are two other competing mechanisms through

which importing possibly increases the demand for skilled labor in developing countries.

First, if importing leads to skill-upgrading through a capital-skill complementarity mech-

anism, rather than the framework posited in Section 2, our estimates may be biased if we

do not explicitly address this additional channel in our benchmark specifications. Second,

Feenstra and Hanson (1996, 1997) present a model with a continuum of goods where the

most skill-intensive goods in developing countries correspond to the least skill-intensive

goods in developed countries. Trade liberalization induces the most skill-intensive goods

in developing countries to be exported to developed countries, leading to an increase in

the demand for skilled labor in developing countries.

We first extend our model in Section 2 to include capital-skill complementarity by con-

sidering the following production function: f(K,M,Ls, Lu, A, ϕ) = ϕ(V p)αp(V n)αnMαm ,

where ϕ is the firm’s Hicks-Neutral productivity shock while V j is a CES aggregator given

by V j = [(Aj(Ljs)β(Kj)1−β)1/ρj + (Lju)

1/ρj ]ρj with ρj = σj/(σj − 1) for j = {n, p}. As be-

fore, Aj captures skill-biased technological change as in our benchmark model. However,

in this case, it augments both skilled labor, Ljs, and capital, Kj through the composite

input (Ljs)β(Kj)1−β. Minimizing the firm’s costs, the relative demand for skilled labor

can be written as:LjsLju

=

(βWu

Ws

)σj(Aj)σj−1

(Kj

Ljs

)(σj−1)(1−β)

, (10)

where we again assume that skill-biased technology is potentially a function of the firm’s

import decision as written in equation (2).

There are three issues here which merit comment. First, equation (10) demonstrates

that if capital-skill complementarity is an important mechanism among Indonesian man-

ufacturers our benchmark specification may potentially suffer from omitted variable bias.

Second, the relative demand for skill equation (10) implies that we need to partition

26

Table 9: Capital-Skill Complementary

Production Workers

Sample Full Full Full Non-Imp Non-ImpDependent Var. ln(Lps/L

pu) ln(Lps/L

pu) ln(Lps/L

pu) ln(Lps/L

pu)

(WsL

ps/(WuL

pu +WsL

ps))06

IV OLS IV using IV using IV using IV usingsmall IV set full IV set full IV set full IV set

(1) (2) (3) (4) (5)

Import Status 0.3581*** 3.4897*** 4.3887*** 1.9506** 0.8892***[0.058] [0.741] [1.209] [0.957] [0.269]

Capital-Skill Comp. -0.2103*** 0.1314 -0.3909 -0.3754* 0.0396*[0.014] [0.407] [0.256] [0.211] [0.023]

No. Obs 4,631 4,288 4,285 3,176 3,310

Non-Production Workers

Sample Full Full Full Non-Imp Non-ImpDependent Var. ln(Lns /L

nu) ln(Lns /L

nu) ln(Lns /L

nu) ln(Lns /L

nu) (WsLns /(WuLnu +WsLns ))06

IV OLS IV using IV using IV using IV usingsmall IV set full IV set full IV set full IV set

(6) (7) (8) (9) (10)

Import Status 0.1711** 2.3698*** 2.1137** 1.5567** 0.2212[0.071] [0.646] [0.848] [0.710] [0.160]

Capital-Skill Comp. -0.1053*** -0.1656 -0.0065 0.1416 -0.0159[0.018] [0.301] [0.229] [0.213] [0.015]

No. Obs 2,417 2,244 2,241 1,642 3,310

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISIC

industry dummies are also included. The small IV set includes the estimated cost of shipping goods to an Indonesian port,

the potential for shipping imports by air, and the weight of imported inputs.

27

Table 10: Skill Demand, Export Decisions, and Initial Skill Levels

Dep. Var.: ln(Lps/Lpu)06 Dep. Var.: Export Dummy in 2006

(1) (2) (3) (4) (5) (6) (7) (8)Exp96 = 0 Exp96 = 0 Full Full Exp96 = 0 Exp96 = 0 Full Full

Sample D96 = 0 D96 = 0 D96 = 0 D96 = 0D06 = 0 D06 = 0 D06 = 0

IV export IV export IV export IV export OLS OLS IV import IV importusing using using and import using using

IV or OLS ∆τy small small using small fullIV set IV set full IV set & IV set IV set

Exp. IV setImport 1.1686** 0.1369*** -0.5194 -0.0999 -0.0112

[0.485] [0.026] [0.387] [0.133] [0.122]Export -2.8811 -1.1937 -1.3266** -1.2819**

[3.417] [1.515] [0.516] [0.322]ln(Lps/L

pu)96 0.3864*** 0.3820*** 0.3920*** 0.3767*** 0.0001 0.0025 -0.0023 -0.0033

[0.034] [0.027] [0.021] [0.023] [0.003] [0.004] [0.004] [0.004]dpu,96 0.1825 0.1742 0.7027 0.5688 -0.0055 0.0001 -0.0166 -0.0146

[0.360] [0.283] [0.186] [0.193] [0.008] [0.010] [0.010] [0.010]dps,96 -1.0704*** -1.0340*** -1.0878*** -1.0349*** -0.0018 0.0061 -0.0301*** -0.0294***

[0.123] [0.087] [0.080] [0.082] [0.008] [0.011] [0.011] [0.010]No. Obs. 2,227 2,333 3,747 3,690 3,925 3,696 5,827 5,824

Notes: ***p < 0.01, **p < 0.05, *p < 0.1. Standard deviations are in square brackets. Province dummies and 3-digit ISICindustry dummies are also included

capital into production and non-production components (Kp and Kn). While the data

do not provide a natural decomposition of capital across occupation, our model implies

that a fraction Ljs/(Lps + Lns ) of total capital K will be allocated to production and non-

production activities and, therefore, Kn = KLns/(Lps + Lns ) and Kp = KLps/(L

ps + Lns ).21

Third, it is clear that capital-skill complementarity implies adding one additional variable

to our benchmark empirical specification, the log ratio of capital to total (production and

non-production) skilled labor, ln (K/(Lps + Lns )). Unfortunately, including the variable

ln (K/(Lps + Lns )) on the right-hand side of equation (5) is very likely to induce endo-

geneity bias since skilled labor determines both outcome and explanatory variables. As

such, columns (2)-(5) of Table 9 document IV estimates of the impact of importing on

the demand skilled labor while also instrumenting the endogenous capital-skill control us-

ing lagged (i.e., 1996) values of ln (K/(Lps + Lns )) as an additional instrument along with

interactions of ln (K/(Lps + Lns )) with our benchmark instruments.22

We find that the impact of importing on the demand for skilled production workers

is nearly unaffected by controlling for capital-skill complementarity; the point estimates

21A full solution is available in Appendix D.22When using the ‘small’ instrument set, lagged ln (K/(Lp

s + Lns )) is interacted with each instrument

in the small instrument set. Analogously, ln (K/(Lps + Ln

s )) is interacted with all instruments when usingthe full instrument set.

28

on the import status variable are of a similar magnitude and significance as the bench-

mark regressions in Table 4. Furthermore, we find that the capital-skill complementarity

control, ln (K/(Lps + Lns )), is not statistically significant across all IV regressions.

We next examine the following two hypotheses that are broadly consistent with the

implications of Feenstra and Hanson’s model. The first hypothesis is that a plant that

starts exporting after trade liberalization must produce skill-intensive goods so that their

demand for skilled labor is higher than other plants. The second hypothesis is that the

higher the initial level of skills, the more likely it is that the plant starts exporting after

trade liberalization. Table 10 shows that neither of these hypotheses appear to hold in

our Indonesian plant-level data. Columns (1)-(2) of Table 10 report the estimated effect

of exporting on the skill ratio among production workers using the sample of initial non-

exporters that do not import in 1996 or 2006 when we instrument exporting, while columns

(3)-(4) report the same estimate when we instrument both importing and exporting using

the full sample. We find no evidence that exporting increases the demand for skilled

production workers conditional on the history of import status. Further, by regressing an

export dummy in 2006 on the log of the skill ratio among production workers in 1996,

columns (5)-(8) of Table 10 show that the initial level of skill intensity among production

workers does not have any predictive power for export status in 2006. These results

indicate that the mechanisms suggested by Feenstra and Hanson model do not appear

relevant for explaining the observed skill upgrading in Indonesia.

4.3 The Demand for Skill: Marginal Treatment Effects

4.3.1 Propensity Scores and the Skill Demand Equation (9)

For brevity, we focus on the estimates of import decisions and the skill demand equation

(9) for the sample of production workers. We estimate the import probabilities for each

plant using a logit specification where we include the interaction terms between instru-

ments and the lagged value of the outcome variable in 1996 as well as dummy variables for

plants that did not hire any skilled or unskilled workers in 1996 as additional explanatory

variables.

Figure 3 plots the distribution of estimated propensity scores for importing and non-

importing plants in both the full sample and the subsample of initial non-importers.23 It

23Table E.2 in Appendix E reports the estimate of the coefficient and marginal derivative for eachvariable with bootstrapped standard errors for the full sample and the sample of initial non-importers,where we find that transport costs and air shares are always a strong predictor of importing.

29

Figure 3: Support of Estimated Propensity Scores

p0 0.2 0.4 0.6 0.8 1 1.2

f(p)

0

0.05

0.1

0.15

0.2

0.25

0.3Density of Import Probabilities for log(L

sp/L

up)

D=0D=1

(a) Full Sample

p0 0.2 0.4 0.6 0.8 1 1.2

f(p)

0

0.1

0.2

0.3

0.4

0.5

0.6Density of Import Probabilities for log(L

sp/L

up)

D=0D=1

(b) Subsample of Initial Non-Importers

is evident that the common support of the propensity scores across importing and non-

importing plants does not span the full unit interval. For this reason, we restrict our

computation of treatment effects to the region where there is significant overlap between

the propensity scores of non-importing and importing plants as reported in the second to

last row of Table 12; specifically, treatment effects are computed over the region with the

minimum and maximum values given by the 1st percentile and the 99th percentile values

of estimated propensity scores for which we have common support, respectively. Because

there are very few non-importing plants with propensity scores beyond the upper bound

of this range, it is difficult to apply nonparametric methods and confidently estimate the

MTE outside of this range.

Table 11 reports the estimates of the parameter γ and δ in the skill demand equation

(9) using the sample of plants for which the outcome variable is measurable and for which

the estimated propensity scores are on the estimated common support. When we use the

skilled worker’s wage share as the outcome variable, the coefficient of the interaction term

between the lagged value of the log of the skill ratio and the propensity score is negative

and significant as reported in the last two columns of Table 11. One possible interpretation

is that plants with high initial skill ratios may have already adopted relatively skill-biased

technology and, as a result, further adoption of foreign technology induced by importing

may not substantially increase their demand for skilled workers. The estimates of the

other explanatory variables are similar to those of the IV regressions in Table 4.

30

Table 11: Estimates of Skill Demand Equation (9) for the Sample of Production Workers

Dependent Var. ln(Lps/Lpu)06 (WsL

ps/(WuL

pu +WsL

ps))06

Sample Full D96 = 0 Full D96 = 0Export -0.3224 [0.0960] -0.2504 [0.1219] -0.0482 [0.0216] -0.0296 [0.0281]Capital 0.0564 [0.0213] 0.0676 [0.0281] 0.0210 [0.0048] 0.0234 [0.0067]Hicks-neutral ϕ -0.2874 [0.0492] -0.3092 [0.0639] -0.0094 [0.0098] -0.0183 [0.0129]Foreign -0.2043 [0.1331] -0.1080 [0.1866] -0.0484 [0.0268] -0.0444 [0.0386]R&D -0.0567 [0.0899] -0.0791 [0.1214] 0.0163 [0.0175] 0.0230 [0.0251]Training 0.1098 [0.0668] 0.0804 [0.0773] 0.0423 [0.0139] 0.0410 [0.0189]log(Ws/Wu)06 -0.1148 [0.1869] -0.1649 [0.2174]log(Ws/Wu)96 -0.3773 [0.2153] -0.6412 [0.2502]log(Lps/L

pu)96 0.3702 [0.0311] 0.3911 [0.0333]

dpu,96 -0.2389 [0.3237] -0.2388 [0.3764]

dps,96 -0.7920 [0.1102] -1.0296 [0.1079]

log(Lps/Lpu)96 × P (Z) -0.0202 [0.0870] -0.3133 [0.2048]

dpu,96 × P (Z) 1.6171 [0.6708] 1.4667 [1.5448]

dps,96 × P (Z) -1.1809 [0.7134] -1.1037 [1.2347]

(WsLps/(WuL

pu +WsL

ps))96 0.0344 [0.0170] 0.0358 [0.0206]

(WsLps/(WuL

pu +WsL

ps))96 × P (Z) -0.4794 [0.0688] -0.4727 [0.1902]

No. Obs. 4094 2911 5564 4097

Notes: The bootstrap standard errors are in square brackets. Province dummies and 3-digit ISIC industry dummies are

also included.

4.3.2 Treatment Effects for Production and Non-production Workers

Figure 4 plots the relationships between the MTE for production workers or non-production

workers and UD along with 90 percent (equal-tailed) bootstrap confidence bands, where

the import decision model is estimated for each bootstrap sample so that the first stage

estimation error is taken into account. As shown in Figure 4(a)(b), the estimated MTE

curve for production workers is well above zero for small values of UD and is downward

sloping in both the full sample and the sample of initial non-importers. These findings

provide evidence that plants self-selected into importing based on their plant-specific, un-

observed component of net benefit of importing—if plants choose to import when their

propensity scores estimated in terms of observables are low, then their unobserved ability

to adopt the skilled biased technology must be high. The results are mixed for non-

production workers. For the full sample, the estimated MTE curve for non-production

workers is above zero and downward sloping for small values of UD as shown in Figure

4(c); in contrast, when we restrict the sample to initial non-importers in Figure 4(d), the

estimated MTE curve for non-production workers is not significantly different from zero

across all values of UD.

Table 12 reports the estimates of various summary measures of the impact of im-

porting on skill demand: the ATE, the TT, the TUT, and policy relevant treatment

31

Figure 4: Estimated MTE for Production Workers and Non-Production Workers

UD

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

MT

E

-1

0

1

2

3

4

5

6

7MTE(p) for log(L

sp/L

up) with 90 percent confidence band

(a) Production, Full Sample

UD

0 0.1 0.2 0.3 0.4 0.5 0.6

MT

E

-5

0

5

10

15

20MTE(p) for log(L

sp/L

up) with 90 percent confidence band

(b) Production, Initial Non-Importers

UD

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9

MT

E

0.5

1

1.5

2

2.5

3

3.5

4

4.5

5MTE(p) for log(L

sn/L

un) with 90 percent confidence band

(c) Non-Production, Full Sample

UD

0 0.1 0.2 0.3 0.4 0.5 0.6

MT

E

-4

-3

-2

-1

0

1

2

3

4

5

6MTE(p) for log(L

sn/L

un) with 90 percent confidence band

(d) Non-Production, Initial Non-Importers

32

effects (the MPRTEs and the PRTEs), where these treatment effects are computed as

the weighted averages of the MTE with weights that integrate to one in the restricted

support reported in the second to last row of Table 12. Appendix B discusses the details

of our estimation procedure and reports the estimated weights for computing different

treatment parameters.

The first four columns of Table 12 report the estimated treatment effects for production

workers across different samples and outcome variables. Bootstrap standard errors and

the 90 percent equal-tailed bootstrap confidence interval are reported in square brackets

and in parentheses, respectively. The ATE, the TT, and the TUT for production work-

ers are estimated to be positive and statistically significant, indicating that importing

increases the demand for educated production workers across different groups of plants.

Furthermore, the TT is estimated to be substantially larger than the ATE which, in turn,

is larger than the TUT. This is potentially indicative of substantial unobserved hetero-

geneity in the effect of importing on skill demand across plants and that plants with

greater returns from importing self-select into importing. While plants that were induced

to import witnessed large increases in the demand for skilled production workers, the

impact of importing on the demand for skilled production workers is substantially smaller

for plants that chose not to import.

33

Table 12: Treatment Effects of Importing on Skill Demand

Production Non-ProductionDep. Var. ln(Lps/L

pu)06 (WsL

ps/(WuL

pu +WsL

ps))06 ln(Lns /L

nu)06 (WsLns /(WuLnu +WsLns ))06

(1) (2) (3) (4) (5) (6) (7) (8)Sample Full D96 = 0 Full D96 = 0 Full D96 = 0 Full D96 = 0ATE 1.9084 2.5591 0.5321 0.7719 1.9245 1.1528 0.3061 0.1267

[0.5119](a) [0.8913] [0.1175] [0.2111] [0.5179] [1.8284] [0.1065] [0.2236]

(1.1197,2.7895)(b) (1.8899,4.7184) (0.3512,0.7438) (0.5733,1.2673) (1.1112,2.9097) (-0.4970,2.7675) (0.1574,0.5076) (-0.1937,0.5345)TT 3.497 7.1728 1.3133 2.9557 2.4924 1.5088 2.024 2.0108

[0.8978] [2.7664] [0.2766] [0.8089] [0.7826] [2.2003] [0.5817] [1.2990](2.1535,5.2215) (4.8884,13.8592) (0.9131,1.8221) (2.3392,5.0455) (1.2513,3.8002) (-2.2790,4.9247) (1.3529,3.3455) (0.4224,4.6746)

TUT 1.5452 2.0999 0.3843 0.5741 1.8294 1.1196 0.1084 -0.0005[0.4991] [0.7991] [0.1024] [0.1702] [0.5379] [2.0998] [0.0813] [0.1744]

(0.7278,2.3760) (1.6360,4.1147) (0.2316,0.5691) (0.3978,0.9644) (1.0644,2.8776) (-0.3431,2.5770) (-0.0366,0.2320) (-0.2766,0.3254)MPRTE 2.9776 5.9734 1.0821 2.4399 2.3229 1.4587 1.0587 1.2223(P ∗α = P + α) [0.7371] [2.2428] [0.2187] [0.6516] [0.6614] [1.9534] [0.2918] [0.8978]

(1.8817,4.3648) (4.1359,11.1331) (0.7578,1.4849) (1.9162,4.0861) (1.2831,3.4162) (-1.7980,4.4199) (0.6641,1.6441) (0.0791,3.0284)MPRTE 2.9113 5.7281 1.051 2.3246 2.2876 1.4393 0.9074 1.0245(Zk∗α = Zk + α) [0.7164] [2.0978] [0.2137] [0.6241] [0.6325] [1.8375] [0.2562] [0.7926]

(1.8403,4.2442) (4.0640,10.6532) (0.7446,1.4426) (1.8298,3.8859) (1.2872,3.3172) (-1.6414,4.2855) (0.5320,1.3996) (-0.0022,2.6159)PRTE 3.0723 6.2576 1.1203 2.5796 2.3437 1.4705 1.3001 1.5049(TC∗ = 0.9TC) [0.7735] [2.5735] [0.2316] [0.6959] [0.6812] [2.0019] [0.3816] [1.0347]

(1.9347,4.5347) (4.4872,11.8936) (0.7730,1.5367) (2.0163,4.3582) (1.2026,3.4337) (-1.8284,4.5247) (0.7796,2.0659) (0.1700,3.5216)

Support(c) [0.0141,0.8697] [0.0043,0.5518] [0.0128,0.8570] [0.0041,0.5483] [0.0090,0.8769] [0.0049,0.5709] [0.0118,0.8652] [0.0047,0.5586]

No. of Obs.(d) 4094 2911 5564 4097 2174 1461 5624 4058

Notes: (a) The bootstrap standard errors are in square brackets. (b) The bootstrap equal-tailed 90 percent confidence intervals are in parentheses. (c) The minimum

and the maximum values of support over which treatment effects are computed; various treatment effects are computed by restricting the weights to integrate to one in

the restricted support, for which minimum and maximum values are determined by the 1st percentile and the 99th percentile of observations in the common support,

respectively. (d) The sample size for estimating the MTE curve.

34

For non-production workers, as the last four columns of Table 12 show, we observe

similar patterns for the ordering among the TT, the ATE, and the TUT across samples

and measures of skill demand. However, the results from the subsample of initial non-

importers reported in columns (6) and (8) indicate that the ATE, the TT, and the TUT

might not be significantly different from zero.

Table 13 examines the robustness of our results using different samples, specifications,

and estimation methods, where we focus on the log of the skill ratio for production workers

as the outcome variable and always use the full sample. In column (1), we estimate the

treatment effects using the subsample of plants that did not export in both 1996 and

2006 to control for the history of export status. Column (2) reports the estimates of

treatment effects when we use conventional TFP in place of our estimated Hicks-neutral

productivity. In column (3), we estimate the partial linear model (9) where we use a sieve

estimator based on the 4th order polynomials in P (Z) instead of the local polynomial

estimator. Column (4) considers a specification of the skill demand equation (9) with no

interactions with Z so that X is empty, while column (5) estimates treatment effects over

the estimated common support instead of the subset of common support defined by the

1st percentile and the 99th percentile of observations that are on the common support.24

The estimates of the ATE, the TT, and the TUT in columns (1)-(5) of Table 13 are

significantly positive and exhibit the patterns similar to those reported in column (1) of

Table 12.

4.3.3 Policy Experiment

Our IV and MTE estimates confirm that importing has a substantial impact on the

demand for skilled production workers. Nonetheless, it is less clear how large the effect of

further changes in policy related variables would be on the demand for skilled production

workers. To examine this issue, we consider alternative policies that change the probability

of importing but do not affect potential outcomes or the unobservables related to import

decisions, (S0, S1, V ) defined in (6)-(7), and compute the mean effect of going from a

baseline policy to an alternative policy per plant shifted into importing. This treatment

effect is called the Policy Relevant Treatment Effect (PRTE) as proposed by Heckman

and Vytlacil (2005, 2007b). Let P ∗(Z) and P (Z) denote the propensity scores under an

alternative policy and a baseline policy, respectively.

24To estimate the treatment effects reported in Table 13, we use the same specification for the decisionto import as the specification reported in Table E.2 except that, in column (4), we exclude the interactionterms between instruments and the 1996 value of the log of skill ratio from the set of explanatory variables.

35

We consider the alternative policy of reducing the cost of shipping goods to the nearest

port by 10 percent so that P ∗(Z) is set to the propensity score under the alternative

transport cost of TC∗ = 0.9TC. The PRTE under this alternative policy captures the

causal impact that a marginal improvement in roads and infrastructure would have on

the relative demand for skilled workers across plants. Note that this policy change will

have a heterogeneous impact across plants. We compute the estimate of what the PRTE

would be when we restrict the support of the propensity scores to the restricted support

reported in the second to the last row of Table 12.25 We also compute the marginal version

of the PRTE called the Marginal Policy Relevant Treatment Effect (MPRTE) proposed by

Carneiro, Heckman, and Vytlacil (2010). Given a sequence of alternative policies indexed

by a scalar variable α such that limα→0 P∗α(Z) = P (Z), the MPRTE is defined as the

limit of a sequence of PRTEs as α approaches to zero. We consider two policy sequences

as described in Carneiro, Heckman, and Vytlacil (2010): (i) a policy that increases the

probability of importing by α so that P ∗α = P + α and (ii) a policy that shifts one of the

components in Z, say Zk, so that Zkα = Zk + α.

As reported in columns (1)-(4) of the lower panel of Tables 12, the estimates of the

MPRTEs and the PRTE for production workers indicate that the subset of plants that

would be induced to start importing by further policy change would substantially increase

their demand for skilled production workers. These estimates are not sensitive to changes

in specifications, samples, and estimation methods on the whole as shown in the lower

panel of Table 13. In contrast, for non-production workers, the results of the PRTE and

the MPRTEs are mixed as reported in columns (5)-(8) of Tables 12.

4.3.4 Discussion

Consistent with our IV results, we find strong evidence that importing has a large and

significant impact on the relative demand for skilled production workers, but evidence

that importing affects the demand for skill among non-production workers is mixed. To

examine whether skill demand is related to the adoption of technology, we also estimated

the treatment effects of adopting standard production processes, where we replace the

import dummy with a dummy for the adoption of standards while using the sample of

25As discussed in Carneiro, Heckman, and Vytlacil (2010), the PRTE is not identifiable without strongsupport conditions. To compute the estimate of what the PRTE would be on the restricted support, wereplace the value of the propensity scores with the maximum value of the support whenever the valueof the propensity scores under the alternative policy is larger than the maximum value of the restrictedsupport so that all of the propensity scores under the alternative policy lie on the restricted support.

36

Table 13: Robustness Check: Treatment Effects of Importing on Skill Demand for Pro-duction Workers using the Full Sample

Dep. Var. ln(Lps/Lpu)06

(1) (2) (3) (4) (5) (6)Sample of Replace ϕ Use Sieve No Treatment Use Dummy for

Export96 = 0 with in place of Interaction Effects over Standards in placeand Export06 = 0 TFP Local Poly. with Z Common Support of Import Dummy

ATE 2.1239 1.7549 1.9936 2.3962 1.8264 1.9084

[0.7234](a) [0.5354] [0.5614] [0.5176] [0.5438] [0.5097]

(1.3980,3.8018)(b) (1.0142,2.7184) (1.0998,2.9811) (1.4846,3.1857) (1.0034,2.8279) (1.0708,2.8227)TT 4.0061 3.4021 4.2679 3.8102 3.4960 3.4970

[1.1810] [0.8994] [1.2917] [0.9304] [0.8900] [0.8978](3.1581,7.0462) (2.1984,5.2046) (2.1246,6.5274) (2.1182,5.1467) (2.1884,5.2206) (2.1678,5.1627)

TUT 1.7609 1.3723 1.5622 2.0671 1.4691 1.5452[0.7220] [0.5284] [0.5593] [0.4749] [0.5609] [0.4963]

(0.8370,3.2672) (0.5946,2.3059) (0.6766,2.4998) (1.2396,2.8302) (0.6164,2.5000) (0.7211,2.3874)MPRTE 3.5934 2.8694 3.3986 3.3587 2.9654 2.9776(P ∗α = P + α) [1.0396] [0.7430] [0.9333] [0.7731] [0.7303] [0.7369]

(2.8156,6.2185) (1.8441,4.3346) (1.8781,5.0529) (1.9970,4.4794) (1.8763,4.3487) (1.8516,4.3422)MPRTE 3.5113 2.7999 3.2643 3.3130 2.8975 2.9113(Zk∗α = Zk + α) [1.0100] [0.7219] [0.8838] [0.7618] [0.7103] [0.7167]

(2.7737,6.0799) (1.8179,4.2036) (1.8355,4.8174) (1.9811,4.4146) (1.8373,4.2556) (1.8274,4.2436)PRTE 3.6612 2.9463 3.5674 3.4498 3.0587 3.0723(TC∗ = 0.9TC) [1.0598] [0.7693] [1.0159] [0.8001] [0.7692] [0.7663]

(2.8656,6.3607) (1.8701,4.4479) (1.8696,5.3462) (2.0518,4.6240) (1.8910,4.4925) (1.9103,4.4780)

Support(c) [0.0114,0.6751] [0.0142,0.8781] [0.0141,0.8697] [0.0153,0.8726] [0.0124,0.9425] [0.0141,0.8697]

No. of Obs.(d) 2519 4100 4094 4099 4094 4094

Notes: (a) The bootstrap standard errors are in square brackets. (b) The bootstrap equal-tailed 90 percent confidence

intervals are in parentheses. (c) The minimum and the maximum values of support over which treatment effects are

computed; various treatment effects are computed by restricting the weights to integrate to one in the restricted support,

for which minimum and maximum values are determined by the 1st percentile and the 99th percentile of observations in

the common support, respectively. (d) The sample size for estimating the MTE curve.

37

initial non-importers which did not export in 1996 or 2006. As reported in column (6) of

Table 13, the TT is estimated to be substantially larger than the ATE or the TUT while

the estimates for the PRTE and the MPRTEs indicate that further policy change which

promotes the adoption of standards would have a substantial impact on the demand for

skilled production workers.

5 Conclusion

This paper studies the impact that importing foreign materials has on the demand for

educated workers among Indonesian manufacturing plants. We develop a model of het-

erogeneous manufacturing plants where the decision to import may be influenced by the

adoption of skill-biased technology. In our model the degree to which importing induces

skill-biased technological change is potentially heterogenous across plants and unobserv-

able to the researcher. To the extent that importing affects skill-biased productivity we

would expect that it will directly impact mix of skilled and unskilled workers hired by

Indonesian manufacturers.

To estimate the impact of importing on the demand for skilled workers we exploit de-

tailed data from the Indonesian manufacturing survey. Our data documents the education

level of every worker in every manufacturing plant with at least 20 employees. Defining

a skilled worker as one with a high school education, we find that importing greatly in-

creases the demand for skilled production workers among Indonesian importers. We also

document significant evidence that the effect of importing on the demand for skilled pro-

duction workers is heterogeneous across plants. In particular, plants which were induced

to import during our sample period were estimated to be those with generally high returns

from importing. We further find that policies that improve transportation infrastructure

in Indonesia would encourage new plants to start importing and increase the demand for

skilled production workers among new importers.

References

Ackerberg, Daniel A., Kevin Caves, and Garth Frazer (2015) “Identification Properties of

Recent Production Function Estimators,” Econometrica, forthcoming.

Amiti, Mary, and Donald Davis (2012) “Trade, Firms, and Wages: Theory and Evidence,”

Review of Economic Studies, 79(1): 1-36.

38

Amiti, Mary, and Lisa Cameron (2012) “Trade Liberalization and the Wage Skill Premium:

Evidence from Indonesia,” Journal of International Economics, 62(2): 277-287.

Amiti, Mary, and Jozef Konings (2007) “Trade Liberalization, Intermediate Inputs, and Pro-

ductivity: Evidence from Indonesia,” American Economic Review, 97(5):1611-1638.

Attanasio, O., Goldberg P., and N. Pavcnik (2004) “Trade Reforms and Wage Inequality in

Colombia,” Journal of Development Economics 74: 331-366.

Autor, D., L. Katz, and A. Krueger (1998) “Computing Inequality: Have Computers Changed

the Labor Market,” Quarterly Journal of Economics, 113: 1169-1214.

Baily, Martin N., Hulten, Charles, and David Campbell (1992) “Productivity Dynamics in

Manufacturing Plants,” Brookings Papers on Economic Activity: Microeconomics, 67-119.

Berman, E., Bound, J., Grilliches, Z. (1994) “Changes in the Demand for Skilled Labour within

U.S. Manfuacturing: Evidence from the Annual Survey of Manufactures,” Quarterly Journal

of Economics 109: 367-397.

Bernard, Andrew B., Jonathan Eaton, J. Bradford Jensen, and Samuel Kortum (2003) “Plants

and Productivity in International Trade,” American Economic Review, 93(4): 1268-1290.

Bernard, Andrew B. and J. Bradford Jensen (1997) “Exporters, Skill Upgrading, and the Wage

Gap,” Journal of International Economics, 42: 1-25.

Bernard, Andrew B., J. Bradford Jensen, Stephen J. Redding and Peter K. Schott (2007)

“Firms in International Trade,” Journal of Economic Perspectives, 21(3): 105-130.

Biscourp, P. and F. Kramarz (2007) “Employment, Skill Structure and International Trade:

Firm-level Evidence for France,” Journal of International Economics, 72(1): 22-51.

Bloom, N., M. Draca and J. Von Reenan (2011) “Trade Induced Technical Change: The Impact

of Chinese Imports on IT and Innovation,” mimeo, Stanford University.

Burstein, Ariel, Javier Cravino and Jonathan Vogel (2013) “Importing Skill-Biased Technolog-

ical Change,” American Economic Journal: Macroeconomics, 5(2): 32-71.

Burstein, Ariel and Jonathan Vogel (2012) “International Trade, Technology, and the Skill

Premium,” mimeo, UCLA.

Bustos, Paula (2011) “The Impact of Trade on Technology and Skill Upgrading: Evidence from

Argentina,” mimeo, University of Pompeu Fabra.

39

Card, David (2001) “Estimating the Return to Schooling: Progress on some Persistent Econo-

metric Problems” Econometrica 69(5): 1127-1160.

Card, David, John E. DiNardo (2002) “Skill Biased Technological Change and Rising Wage

Inequality: Some Problems and Puzzles,” Journal of Labor Economics, 20: 733-783.

Carneiro, P., J.J. Heckman, and E. Vytlacil (2010) “Evaluating Marginal Policy Changes and

the Average Effect of Treatment for Individuals at the Margin,” Econometrica 78(1): 377-394.

Carneiro, P., J.J. Heckman, and E. Vytlacil (2011) “Estimating Marginal Returns to Educa-

tion,” American Economic Review, 101: 2754-2781.

Chor, D. (2010) “Unpacking Sources of Comparative Advantage: A Quantitative Approach.,”

Journal of International Economics, 82(2): 152-167.

Clerides, Sofronis, Lach, Saul, and James R. Tybout (1998) “Is Learning by Exporting Impor-

tant? Micro-Dynamic Evidence from Colombia, Mexico, and Morocco,” Quarterly Journal of

Economics 113: 903-947.

Cosar, A. Kerem (2013) “Adjusting to Trade Liberalization: Reallocation and Labor Market

Policies,” mimeo, Stockholm School of Economics.

Cosar, A. Kerem and Banu Demir (2015) “Domestic Road Infrastructure and International

Trade: Evidence from Turkey,” mimeo, Stockholm School of Economics.

Costinot, Arnaud and Jonathan Vogel (2010) “Matching and Inequality in the World Econ-

omy,” Journal of Political Economy, 118(4): 747-786.

Davis, Donald R. and James Harrigan (2011) “Good Jobs, Bad Jobs, and Trade Liberalization,”

Journal of International Economics, 84(1), 26-36.

De Loecker, Jan (2011) “Product Differentiation, Multi-Product Firms and Estimating the

Impact of Trade Liberalization on Productivity,” Econometrica, 79(5): 1407-1451.

De Loecker, Jan, Pinelopi Goldberg, Amit Khandelwal and Nina Pavcnik (2012) “Prices,

Markups and Trade Reform,” mimeo, Princeton University.

Doms, Mark, Timothy Dunne, and Kenneth R. Troske (1997) “Workers, Wages, and Technolo-

gies,” Quarterly Journal of Economics, 112: 253-290.

Doraszelski, Ulrich and Jordi Jaumandreu (2013) “R&D and Productivity: Estimating Produc-

tion Functions When Productivity is Endogenous,” Review of Economic Studies, forthcoming.

40

Duflo, Esther (2001) “Schooling and Labor Market Consequences of School Construction in

Indonesia: Evidence from an Unusual Policy Experiment,” American Economic Review, 91(4):

795-813.

Ebenstein, Avraham, Ann Harrison, Margaret McMillan, Shannon Phillips (2014) “Estimat-

ing the Impact of Trade and Offshoring on American Workers using the Current Population

Surveys,” Review of Economics and Statistics, 96(4): 581-595.

Egger, H. and U. Kreikemeier (2009) “Firm Heterogeneity and the Labor Market Effects of

Trade Liberalization,” International Economic Review, 50 (1): 187-216.

Fan, J. and T.H. Yim (2004) “A Crossvalidation Method for Estimating Conditional Densities,”

Biometrika, 91: 819-834.

Felbermayr, G., J. Prat, H. Schmerer (2011) “Globalization and labor market outcomes: Wage

bargaining, search frictions, and firm heterogeneity,” Journal of Economic Theory, 146 (1):

39-73.

Feenstra, Robert and Gordon Hanson (1996) “Globalization, Outsourcing, and Wage Inequal-

ity,” American Economic Review, 86(2): 240-245.

Feenstra, Robert and Gordon Hanson (1997) “Foreign Direct Investment and Relative Wages:

Evidence from Mexico’s Maquiladoras,” Journal of International Economics, 42(3/4): 371-393.

Feenstra, Robert and Gordon Hanson (1999) “The Impact of Outsourcing and High-Technology

Capital on Wages: Estimates for the United States, 1979-1990,” Quarterly Journal of Eco-

nomics, 114 (3): 907-940.

Frazer, Garth (2013) “Imports, Import Sources and Skill Utilization,” mimeo, University of

Toronto.

Frıas, J. A., D. S. Kaplan and E. A. Verhoogen (2009) “Exports and Wage Premia: Evidence

from Mexican Employer-Employee Data,” mimeo, Columbia University.

Gandhi, Amit, Salvador Navarro and David Rivers (2013) “On the Identification of Production

Functions: How Heterogeneous is Productivity,” mimeo, University of Wisconsin.

Gindling, T.H., Robbins, D. (2001) “Patterns and Sources of Changing Wage Inequality in

Chile and Costa Rica During Adjustment,” World Development, 29: 725-745.

Goldberg, P.K. and N. Pavcnik (2005): “Trade Protection and Wages: Evidence from the

Colombian Trade Reforms,” Journal of International Economics, 66(1): 75-105.

41

Goldberg, P.K. and N. Pavcnik (2007) “Distributional Effects of Globalization in Developing

Countries,” Journal of Economic Literature, Vol. XLV, pp. 39–82

Gonsaga, Gustavo, Naercio Menezes Filho, and Christina Terra (2006) “Trade liberalization

and the evolution of skill earnings differentials in Brazil,” Journal of International Economics,

68: 345-367.

Griliches, Zvi (1969) “Capital-Skill Complementarity,” Review of Economics and Statistics,

51: 465-468.

Halpern, Laszlo, Miklos Koren, and Adam Szeidl (2015) “Imports and Productivity,” American

Economic Review, forthcoming.

Harrison, Ann, and Gordon Hanson (1999) “Who Gains from Trade Reform? Some Remaining

Puzzles,” Journal of Development Economics, 59: 125-154.

Heckman, J.J. (1979) “Sample Selection Bias as a Specification Error,” Econometrica 47:

153-161.

Heckman, J.J., and E.J. Vytlacil (1999) “Local Instrumental Variables and latent Variable

models for Identifying and Bounding Treatment Effects” Proceedings of the National Academy

of Sciences 96: 4730-4734.

Heckman, J.J., and E.J. Vytlacil (2005) “Structural Equations, Treatment Effects and Econo-

metric Policy Evaluation,” Econometrica 73(3): 669-738.

Heckman, J.J., and E.J. Vytlacil (2007a) “Econometric Evaluation of Social Programs, Part I:

Causal Models, Structural Models and Econometric Policy Evaluation” in Heckman, J. and E.

Leamer (Eds.), Handbook of Econometrics, vol. 6B. North-Holland, New York:: 4780-4874.

Heckman, J.J., and E.J. Vytlacil (2007b) “Econometric Evaluation of Social Programs, Part

II: Using the Marginal Treatment Effect to Organize Alternative Econometric Estimators to

Evaluate Social Programs, and To Forecast their Effects in New Environments” in Heckman,

J. and E. Leamer (Eds.), Handbook of Econometrics, vol. 6B. North-Holland, New York::

4875-5143.

Helpman, E., O. Itskhoki, and S. Redding (2010) “Inequality and Unemployment in a Global

Economy,” Econometrica, 78(4): 1239-1283.

Hendayana, Sumar, Asep Supriatna and Harun Imansyah (2008) “Indonesia’s Issues and Chal-

lenges on Quality Improvement of Mathematics and Science Education,” Indonesia University

of Education, Indonesia.

42

Hummels, David, Rasmus Jørgensen, Jakob Munch, and Chong Xiang (2014) “The Wage

Effects of Offshoring: Evidence from Danish Matched Worker-Firm Data,” American Economic

Review, 104(6): 1597-1629.

Hummels, David L., and Georg Schaur (2013) “Time as a Trade Barrier,” American Economic

Review, 103(7): 2935-2959.

Imbens, Guido, and Joshua Angrist (1994) “Identification and Estimation of Local Average

Treatment Effects,” Econometrica 62(2): 467-475.

Kam, Ho Wah (2006) “English Language Teaching in East Asia Today: An Overview,” Asia

Pacific Journal of Education, 22(2): 1-22

Kasahara, Hiroyuki and Beverly Lapham (2013) “Productivity and the Decision to Import and

Export: Theory and Evidence,” Journal of International Economics, 89(2), 297-316.

Kasahara, H. and J. Rodrigue (2008) “Does the Use of Imported Intermediates Increase Pro-

ductivity? Plant-level Evidence,” Journal of Development Economics 87: 106-118.

Kijima, Y. “Why did wage inequality increase? Evidence from urban India 1983-99,” Journal

of Development Economics, 81: 97-117.

Koren, M. and M. Csillag (2011) “Machines and machinists: capital-skill complementarity from

an international trade perspective,” CeFiG Working Papers, No. 13.

Krishna, Pravin, Jennifer P. Poole, and Mine Z. Senses (2011) “Wage effects of trade reform

with endogenous worker mobility,” Journal of International Economics, 93(2): 239-252.

Kugler, M. and E. Verhoogen (2009) “Plants and Imported Inputs: New Facts and an Inter-

pretation,” American Economic Review Papers and Proceedings, 99(2):501-507.

Lileeva, Alla, and Daniel Trefler (2010) “Improved Access to Foreign Markets Raises Plant-

Level Production...For Some Plants,” Quarterly Journal of Economics, 125(3):1051-1099.

Levinsohn, James and Amil Petrin (2003) “Estimating Production Functions Using Inputs to

Control for Unobservables,” Review of Economic Studies, 70:317-341.

Martins, Pedro S. and Luca David Opromolla (2009) “Exports, Importes and Wages: Evidence

from Matched Firm-Worker-Product Panels,” IZA Discussion Paper No. 4646.

Matsuyama, Kiminori (2007) “Beyond Icebergs: Towards A Theory of Biased Globalization,”

Review of Economic Studies, 74: 237-253.

43

Muendler, Marc-Andreas (2004) “Trade, Technology, and Productivity: A Study of Brazilian

Manufacturers, 1986-1998,” UCSD, mimeo.

Menezes-Filho, Nae rcio A. and Marc-Andreas Muendler (2011) “Labor reallocation in response

to trade reform,” mimeo, UCSD.

Olley, G. Steven and Ariel Pakes (1996) “The Dynamics of Productivity in the Telecommuni-

cations Equipment Industry,” Econometrica, 65(1); 245-276.

Parro, Fernando (2013) “Capital-Skill Complementarity and the Skill Premium in a Quantita-

tive Model of Trade,” American Economic Journal: Macroeconomics, 5(2): 72-117.

Pavcnik, Nina (2002) “Trade Liberalization, Exit, and Productivity Improvements: Evidence

From Chilean Plants,” Review of Economic Studies, 69(1): 245-276.

Pavcnik, Nina (2003) “What explains skill upgrading in less developed countries?” Journal

of Development Economics 71: 311-328.

Pavcnik, N., A. Blom, P. Goldberg and N. Schady (2004) “Trade liberalization and industry

wage structure: evidence from Brazil,” World Bank Economic Review, 18: 319-344.

Robinson, P. M. (1988) “Root-N-Consistent Semiparametric Regression,” Econometrica, 56(4):

931-954.

Sanchez-Paramo, C. and N. Schady (2003): “Off and Running? Technology, Trade, and the

Rising Demand for Skilled Workers in Latin America,” World Bank Policy Research Working

Paper 3015. Washington, DC: World Bank.

Trefler, Daniel (2004) “The Long and Short of the Canada-U.S. Free Trade Agreement,” Amer-

ican Economic Review, 94(4): 870-895.

Vannoorenberghe, Gonzague (2011) “Trade between symmetric countries, heterogeneous firms,

and the skill premium,” Canadian Journal of Economics, 44(1):148-170.

Verhoogen, Eric (2008) “Trade, Quality Upgrading and Wage Inequality in the Mexican Man-

ufacturing Sector,” Quarterly Journal of Economics, 123(2): 489-530.

Vytlacil, Edward (2002) “Independence, Monotonicity, and Latent Index Models: An Equiva-

lence Result,” Econometrica, 70(1): 331-341.

Wooldridge, Jeffrey (2009) “On Estimating Firm-Level Production Functions Using Proxy

Variables to Control for Unobservables,” Economic Letters, 104:112-114.

Yeaple, Stephen (2005) “A Simple Model of Firm Heterogeneity, International Trade, and

Wages,” Journal of International Economics, 65(1): 1-20.

44

(Not For Publication) Supplementary Appendix to

“Does Importing Intermediates Increase the Demand

for Skilled Workers? Plant-level Evidence from

Indonesia”

Hiroyuki Kasahara∗

Vancouver School of EconomicsUniversity of British Columbia

Yawen LiangVancouver School of EconomicsUniversity of British Columbia

Joel RodrigueDepartment of EconomicsVanderbilt University

A Data Description

This section describes the construction of the variables used for the empirical analysis. Specifically, itincludes the description of variables coming from manufacturing survey data, household survey data andthe construction of instrumental variables.

A.1 Manufacturing Plant Data

Our plant level data comes from the Indonesian manufacturing census (Large and Medium IndustrialStatistics) in years 1994-1996 and 2004-2007. This survey data covers all manufacturing plants in In-donesia with at least 20 employees. Key variables used in our study are described below.

A.1.1 Labor

For each plant, the survey records the education levels of all production and non-production workers.This dimension of the data allows us to compute the number of skilled and unskilled workers in eachoccupation category. As discussed in the main text we define a skilled worker as a worker with a highschool diploma or more. Using this threshold we count the number of skilled and unskilled workers foreach production category and each plant.

A.1.2 Intermediate Goods and Capital

In order to estimate plant-specific productivity, we also need the intermediate goods and capital used forproduction. Intermediate goods includes imported raw materials, domestically purchased raw materialsand expenditures on energy. The wholesale price index for manufactured goods is used to convert nominalvalues into real values.

∗Address for correspondence: Hiroyuki Kasahara, Vancouver School of Economics, University of BritishColumbia, 997-1873 East Mall, Vancouver, BC, V6T 1Z1 Canada.

1

We compute the real value of capital at the beginning of year t as

Kit = buildingit/Pbuildt +machineit/P

macht + vehicleit × 100/P vehict + (rentit/0.1)/P rentt ,

where buildingit, machineit, and vehicleit are the nominal value of buildings, machines, and vehiclesat the beginning of year t; rentit is equal to the reported value of rental payments for buildings andmachines, where we divide the rental value by the depreciation rate (10 percent) to get the appropriatecapitalized value. The capital price indices are obtained from Badan Pusat Statistik (BPS).1 Since rentis only paid for buildings and machines, we compute price index for rented capital as

P rentt =

∑i buildingit∑

i(buildingit +machineit)× P buildt +

∑imachineit∑

i(buildingit +machineit)× Pmacht .

When the capital values are not reported in 1996 or 2006, we use the reported values of capital in1994, 1995 and 1997 for constructing the 1996 capital value, and similarly, the reported values of cap-ital in 2004, 2005 and 2007 for constructing the 2006 capital value by assuming Kit = 0.9Kit−1 +Investmentit−1 with Investmentit = Investmentbuildingsit + Investmentmachinesit + Investmentvehiclesit ,

where Investmentbuildingit , Investmentmachinesit , and Investmentvehiclesit are the real values of net invest-ment for buildings, machines, and vehicles in year t.

Some plants do not report capital values in any year between 2004 and 2007. For those plants, weimpute the values of capital as follows. First, using the plant observations in 2005 for which capitalvalues are constructed from the data between 2004 and 2007, we run the OLS regression logKi,2005 =X ′i,2005α + εi,2005, where Ki,2005 is the beginning-of-period capital in 2005; Xit−1 includes a constant,the ratio of investment to capital, the log of production workers, the log of non-production workers, thelog of output, the log of intermediate goods, an import dummy, province dummies, industry dummies,plant age, plant age squared, a dummy variable for positive investment, a dummy variable for no hiringof production workers, and a dummy variable for no hiring of non-production workers. Then, given theOLS estimate of α, α, we compute the imputed value of capital at the beginning of year 2006 for plantswith missing capital values as Kimpute

i,2006 = 0.9 exp(X ′i,2005α) + Investmenti,2005. For the sample of initialnon-importers, we use the imputed values of capital for 11 percent of observations. For plants withmissing capital values in 1996, we similarly construct the imputed value of capital at the beginning ofyear 1996 using 1995 data.

A.1.3 Other Variables

Other plant information contained in the data includes the percentage of foreign ownership, total expenseson research and development (R&D), and total expenses on training. Dummies variables for foreignownership, R&D and training are defined as whether the above mentioned variables are greater thanzero.

A.2 Regional Data

A.2.1 Wages

We use the household survey data (SAKERNAS—Indonesian Labour Force Survey) to construct localmarket wages for skilled and unskilled workers for each regency (kabupaten/kota) in Indonesia. The

1Specifically, we use the price indices for construction goods, imported and domestic machines, and vehicles.The imported and domestic machines price indices are weighted according to the input-output table for manu-factured goods to get one price index for machines. The building price index covers the period 1996-2006 and isextended to 2007; machine and vehicle price data only covers 1998 to 2005 and is extended to the period 1996-2007. The extension from 1998 to 1996 relies on the wholesale price of capital goods which is available during the1992-1999 period. The GDP deflator for construction goods, machines and vehicles is used to extend the originalprice index to 2007.

2

basic measure of a wage is the average wage of workers who are 25-35 years old at different skill levels(more or less than high-school education) in the regency. However, the relative wage computed using thisbasic measure may not measure the true skill premium since unskilled workers will generally have hadmore work experience than skilled workers within same age range. To solve this problem, we use Mincerregressions to get the return to skill. Specifically, for each regency, we regress the individual level logwage over experience, experience squared and a skill dummy and use the coefficient on the skill dummyas the log of the wage skill premium in that regency. The plant-level wage share of skilled workers in the

total wage bills for occupation j ∈ {p, n} is computed asWsL

js

WsLjs+WuL

ju

=(Ws/Wu)L

js

(Ws/Wu)Ljs+L

ju

, where Ws/Wu is

the estimated wage skill premium for the regency in which a plant is located.

A.2.2 Distance to Port

To form this instrumental variable, we use the location information of individual plants. Indonesia iscomprised of 33 provinces which are administratively subdivided in to 429 regencies in our data. Thedataset includes the location of the surveyed plants down to the regency level. Due to the detailedadministrative divisions, the variation contained in the plant location data is informative. Among all portsin Indonesia, 2 can be considered large, 16 medium and all others remaining are either small or very small.2

The 18 large or medium sized ports are chosen to be the main destinations for our constructed measureof transportation costs. Specifically, given these destinations, and taking the geographical features ofIndonesia into consideration, we compute the least-cost path from the center of every regency to itsnearest port by ArcGIS. The calculation divides the entire country into cells with size 1 km2 where eachcell contains a value representing the average elevation of that area. The travel cost of each cell dependson the slope from the cell to its adjacent cells and whether the cell locates on land or on sea. ArcGISdetermines the optimal route for each cell by finding the least-accumulative-cost path to its nearestmajor port. The transportation cost for a regency is approximated by the the accumulative cost alongthe optimal route from the center cell of the regency. For each plant, the proxy for its transportationcost is the transportation cost of the regency in which the plant is located. Details about the process ofcomputing this cost measure are described in the following paragraphs.

Three types of data are used in ArcGIS to generate the transportation cost: raster data (R), pointdata (P) and table data (T). Raster data consists of a matrix of cells (pixels) organized into a grid whereeach cell contains a value representing information. In our data, each cell represents a 1 km2 squarein the real world. Point data contains information for specific points. Each point is composed of onecoordinate pair representing its location on the earth. Table data is used to store the attributes (e.g.names, locations, temperatures, etc.) of features.

There are three main steps for computing the transportation cost. First, generate the cost rasterfor Indonesia which defines the cost to move planimetrically through each cell according to geographicalfeatures. Second, given a cost raster and the main ports as destination points, the “Cost Distance”tool generates the raster data in which the least accumulated cost distance for each cell to its nearestdestination is calculated. Lastly, to get the measure of the transportation cost for each regency, we extractthe cost distance value for the cells located in the center of the regencies from the raster data obtainedfrom second step. Figure A.1 displays the process of this calculation. The ellipses in the flowchartrepresent data while the round-cornered squares represent tools.

Step 1. The travel cost of each cell depends on the slope from the cell to its adjacent cells andwhether the cell is located on land or sea. “Elevation-full” is the Indonesia elevation data, the value ofa cell in this raster data indicates the average elevation in the 1 km2. Cells in the sea take a value ofzero. The “SLOPE” tool generates the slope layer “Elevation Slope”, in which a cell value indicates themaximum rate of change between the cell and its neighbors. A road which traverses less steep slopes ispreferable. We reclassify the slope layer, slicing the values into 10 equal intervals. A value of 10 is assignedto the most costly slopes (steepest) and 1 is assigned to the least costly slope (flattest), values in betweenare ranked linearly. “Reclass Slope” is the raster data after re-classification. Each cell value between 1

2Source of port information: World Port Source.

3

and 10 indicates the difficulty of traveling over it. One problem with this surface is that traveling acrossthe sea is considered costless since the elevation is zero (and so are the slopes) everywhere on the sea.To solve this problem, another layer “Sea” is created. The “Sea” raster assigns value 0 for land and1 for sea. The last step for generating the cost raster overlays the rasters “Reclass Slope” and “Sea”using a common measurement scale and weights 50 percent on each layer. Specifically, scale values ofthe “Reclass Slope” layer are unchanged (10 for steepest and 1 for flattest), and scale values for “Sea”layer are set to be 1 for land (low cost) and 10 for sea (high cost), thus, the cost of travelling over cell iis Costi = 0.5× ReclassSlopei + 0.5× 10Seai . Putting all the cells on map forms the raster data “CostSurf”.

Step 2. Given the 18 main ports (“Main Ports”) as destinations, the “COST DISTANCE” toolcalculates the accumulated distance from each cell to its nearest destination along the optimal path,using the “Cost Surf” data obtained in step 1 to measure the cost of passing cells. The resulting rasterdata “Cost Dist” reports the transportation cost of all the cells.

Step 3. We extract the values of the cells located in the center of administrative regencies from thetransportation cost map “Cost Dist” using the tool “EXTRACT VALUES TO POINTS.”

Figure A.1: Process of Measuring Transportation Cost

Notes: This figure displays the process of calculating the transportation cost for the regencies in Indonesia usingArcGIS. The ellipses in the flowchart represent data and the round-cornered squares represent tools.

4

Figure A.2: Change in Tariffs, 1991-2000, Relative to 1991 Level

−.5

0.5

11.

5C

hang

e in

Tar

iffs

1991

−20

00

0 .2 .4 .6 .8Tariff in 1991

Notes: Tariffs fell over the sample period in all industries with the exception of the liquors and wine industries(ISIC codes 31310, 31320) and rice milling industries (ISIC codes 31161, 31169).

A.2.3 Tariffs

The plant-level manufacturing data are matched with detailed industry-level tariff data from Amitiand Konings (2007). Specifically, each plant is assigned a 5-digit code in each year and is matchedto output and input tariff rates with the same 5-digit code in the same year. Figure A.2 demonstratesthat output tariffs have fallen across most industries in Indonesia over the 1991-2001 period and thatthere is substantial variation in the initial tariff levels and the subsequent fall across 5-digit industriesover the following decade.

A.2.4 Import Heaviness and Airshare

This section describes our measures of the heaviness of imported inputs (import weight) and the fractionto of imported inputs shipped by air (import airshare) as described in the main text. We first create proxyvariables for transport intensity at HS6 level for Indonesian imports using data on US and EU importsto Indonesia by mode of transportation for the year 2006. Detailed data for U.S. exports by commod-ity and transport mode are published by the US census at http://www.census.gov/foreign-trade/

reference/products/layouts/imdb.html#imp_detl. Similar data for EU exports is taken from theEU International Trade Database ComExt which is published at http://epp.eurostat.ec.europa.eu/newxtweb/. The underlying data set for our EU instruments is collected in the dataset named ‘EXTRAEU Trade Since 2000 By Mode of Transport (HS6) (DS-043328).’ We then follow Cosar and Demir (2015)to construct the heaviness and fraction of imports shipped to Indonesia at each HS6 commodity code forboth the EU and US series separately.

To create the measures of imported input heaviness and airshare we need to map the HS6 measuresabove to the import input-output matrix produced by BPS Indonesia. A key intermediate step in thisprocess is linking the HS6 commodity codes to ISIC 3.0 industry classification in order to create industry-level import variables. To complete this task we use the correspondence table ‘2002 NAICS to ISIC 3.1’as published by the U.S. Census (https://www.census.gov/cgi-bin/sssd/naics) and the correspon-dence table ‘isic31 to isic3’ from the United Nations Stats Division published online at http://unstats.un.org/unsd/cr/registry/regot.asp?Lg=1. For robustness, we repeat this concordance using the cor-respondence tables ‘2002 NAICS to ISIC 4’, ‘isic4 to isic31’ and ‘isic31 to isic3’. These produce similarresults.

5

Last, we use the import input-output Table produced by BPS Indonesia (2000) to construct theimported input measures of heaviness and airshare. The input-output matrix provided by BPS Indonesiaallows us to determine the share of import expenditures in each sector. Specifically, we subtract totaldomestic expenditures in any given sector from total expenditures in the same sector. For each sectorwe can then straightforwardly compute the share of total expenditures on imports from each individualsector.

The input-output tables also provide a concordance between ISIC 3.0 classifications and IndonesianIO sectors. The IO tables are comprised 175 distinct ‘sectors’ which typically aggregate several ISIC 3.0classifications. To determine the sectoral heaviness or airshare, we assign equal shares to all ISIC 3.0classifications assigned to the same sector. As described in the main text, we then use the sectoral importexpenditures shares to construct a measure imported input weight and airshare.

B Estimating MTE and Treatment Effects

We estimate γ, δ, and K(p) by a partially linear regression of S on X and P (Z) (Robinson, 1988) asfollows.

Step 1: We estimate P (Z) using a logit specification as described in the main text. Denote the estimatedvalue by “hat” notation so that P (Z) denotes the estimate of P (Z).

Step 2: Using the subsample of observations for which the outcome variable is measurable and for whichestimated propensity scores P (Zi)’s are on the estimated common support, we estimate E[S|P (Z)],E[X|P (Z)], and E[X|P (Z)] by local linear regressions of S, X, and X on P (Z), respectively, wherewe use a normal kernel and choose their bandwidths by “leave-one-out” cross-validation.

Step 3: By regressing S− E[S|P (Z)] on X − E[X|P (Z)] and P (Z)(X − E[X|P (Z)]) without an intercept,we obtain the estimate of γ and θ.

Step 4: We estimate K(P (Z)) and K ′(P (Z)) by using a local quadratic regression of S −X ′γ − P (Z)X ′θon P (Z), where we use cross-validation to choose the bandwidth for the local quadratic regression.

To avoid numerical singularity, all continuous variables in Z, X, and X are standardized by subtractingtheir means and then dividing by their sample standard deviations while all dummy variables are trans-formed into {−1, 1}. Table B.1 reports the bandwidth choices using the standardized variables for Step2 and Step 4. We set the maximum value of the bandwidth to one-half of the length of the commonsupport of P (Z|D = 0) and P (Z|D = 1).

In column (3) of Table 13, we use a sieve estimator to estimate the partial linear regression. Specif-ically, we estimate E[S|P (Z)], E[X|P (Z)], and E[X|P (Z)] in Step 2 by regressing S, X, and X on thefourth order polynomials of P (Z) while we estimate K(P (Z)) and K ′(P (Z)) by regressing S − X ′γ −P (Z)X ′θ on the fourth order of polynomials in P (Z).

As Heckman and Vytlacil (2005, 2007a, 2007b) and Carneiro, Heckman, and Vytlacil (2010) show,various treatment effects conditional on X can be expressed as weighted averages of the MTE as follows:

ATE(x) =

∫ 1

0

∆MTE(x, p)dp, TT (x) =

∫ 1

0

∆MTE(x, p)hTT (x, p)dp,

TUT (x) =

∫ 1

0

∆MTE(x, p)hTUT (x, p)dp, PRTE(x) =

∫ 1

0

∆MTE(x, p)hPRTE(x, p)dp,

MPRTE(x) =

∫ 1

0

∆MTE(x, p)hPRTE(x, p)dp,

(11)

6

where

hTT (x, p) =1− FP (p|X = x)

E(P |X = x), hTUT (x, p) =

FP (p|X = x)

E(1− P |X = x),

hPRTE(x, p) =FP∗(p|X = x)− FP (p|X = x)

E(P |X = x)− E(P ∗|X = x),

hMPRTE(x, p) = limα→0

FP∗α (p|X = x)− FP (p|X = x)

E(P |X = x)− E(P ∗α|X = x)=

(∂/∂α)FP∗α (p|X = x)|α=0∫(∂/∂α)FP∗α (p|X = x)|α=0dp

.

(12)

FP (·|X = x) and FP∗(·|X = x) are the cumulative distributions of P and P ∗, respectively, conditionalon X = x, where P ∗ is the probability of importing under an alternative policy.

Treatment effects can be computed by integrating conditional treatment effects in (11) using theappropriate distribution of X. Because X is high dimensional, however, it is not computationally feasibleto estimate the conditional density function of P given X. For this reason, exploiting the fact thatfp(P |X) = fp(P |X ′θ) implies E[log(P/(1− P ))|X] = E[log(P/(1− P ))|X ′θ], we regress log(P /(1− P ))

on X and obtain a single index of X, X ′θ. The conditional density function of P given X ′θ, denoted byfP (p|x′θ), is estimated by the ratio of the joint density of P and X ′θ to the marginal density of X ′θ using‘double-kernel’ local linear regression, where we choose the bandwidth by the cross-validation followingthe suggestion of Fan and Yim (2004).

We compute weights hTT (x′θ, p), hTUT (x′θ, p), hPRTE(x′θ, p), and hMPRTE(x′θ, p) as hTT (x, p),hTUT (x, p), hPRTE(x, p), and hMPRTE(x′θ, p) in the formula (12) but using FP (p|X ′θ = x′θ) =

∫ p0fP (u|X ′θ =

x′θ)du in place of FP (p|X = x). To apply (11) to compute treatment effects conditioning on the singleindex X ′θ, we evaluate the MTE at X ′θ = x′θ instead of X = x. To do so, we estimate E[X ′δ|X ′θ] bylocal linear regression and define the MTE at X ′θ = x′θ as ∆MTE(x′θ, p) = E[x′δ|X ′θ = x′θ] + K ′(p).Integrating ∆MTE(x′θ, p) using weights hTT (x′θ, p), hTUT (x′θ, p), hPRTE(x′θ, p) , and hMPRTE(x′θ, p)gives our estimates of the TT (x′θ), TUT (x′θ), PRTE(x′θ), and MPRTE(x′θ). To obtain the uncon-ditional version of treatment effects, we integrate X ′θ from TT (X ′θ), TUT (X ′θ), PRTE(X ′θ), andMPRTE(X ′θ) using the marginal distribution of X ′θ, denoted by fX′θ(x

′θ), which is estimated by locallinear regression. The last three rows of Table B.1 report the bandwidth choices associated with esti-mating fP (p|x′θ) and fX′θ(x

′θ). Figure B.3 shows estimated weights for ATE, TT, TUT, MPRTEs, andPRTE when dependent variable is ln(Lps/L

pu).

Finally, because the full support condition is violated, we report estimates of ATE, TT, TUT, PRTE,

and MPRTE when we restrict the weights to integrate to one in the restricted support of the MTE as

described in the main text. As discussed in Heckman and Vytlacil (2005) and Carneiro, Heckman and

Vytlacil (2010), the PRTE cannot be identified without strong support conditions. We compute the

estimate of what the PRTE would be when we restrict the support of P and P ∗ to the restricted support

for which minimum and maximum values are given by the 1st and the 99th percentiles of the common

support. When the value of P ∗ is above the maximum value of the support, the maximum value of P ∗

is set to the maximum value of the restricted support.

We use 500 bootstrap replications to construct equal-tailed bootstrap confidence bands for

∆MTE(x′θ, p) and the standard errors for treatment effects. In each bootstrap iteration we

re-estimate P (Z) so all standard errors account for the fact that P (Z) is estimated.

7

Table B.1: Bandwidth Choices by Cross-validation

Dep. Var. ln(Lps/Lpu)06 (

WsLps

WuLpu+WsL

ps

)06

(1) (2) (3) (4)Sample Full D96 = 0 Full D96 = 0Step 2: E[S|P ] 0.07 0.03 0.03 0.01

E[Export|P ] 0.15 0.05 0.09 0.05E[Capital|P ] 0.05 0.03 0.03 0.03E[ϕ|P ] 0.05 0.09 0.03 0.05E[Foreign|P ] 0.15 0.39 0.15 0.38E[R&D|P ] 0.11 0.15 0.21 0.17E[Training|P ] 0.05 0.05 0.03 0.03E[ln(Ws/Wu)06|P ] 0.47 0.39E[ln(Ws/Wu)96|P ] 0.11 0.07E[ln(Lps/L

pu)96|P ] 0.47 0.03

E[dpu,96|P ] 0.01 0.39

E[dps,96|P ] 0.03 0.03

E[(WsL

ps

WuLpu+WsL

ps

)96|P ] 0.05 0.01

E[industry/province|P ](a) 0.05 0.03 0.03 0.01

Step 4: E[S −X′γ − P (Z)X′θ|P ] 0.23 0.09 0.15 0.07Bandwidth for P of fP (p|x′θ) 0.005 0.005 0.005 0.005Bandwidth for X′θ of fP (p|x′θ) 0.02 0.01 0.01 0.01Bandwidth for fX′θ(x′θ) 0.03 0.02 0.02 0.02

Notes: Columns (1)-(4) reports the cross-validation bandwidth choices that are used to estimate the treatment effects

reported in columns (1)-(4) of Table 12, respectively. (a) We choose the common bandwidth for industry/province

dummies by minimizing the sum of cross-validation criterion functions over industry/province dummies.

Figure B.3: Weights for ATE, TT, TUT, MPRTEs, and PRTE (Dependent Variables:ln(Lps/L

pu))

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9−0.05

0

0.05

0.1

0.15

0.2

UD

Wei

ght

Weight function for PRTE and MPRTE (outcome variable: log(Lsp/L

up))

ATETTTUTPRTE: (1−α)TCMPRTE: Pα=P+α

MPRTE: Zkα=Zk+α

(a) Production Workers, Full Sample

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4 0.45 0.5−0.05

0

0.05

0.1

0.15

0.2

0.25

0.3

UD

Wei

ght

Weight function for PRTE and MPRTE (outcome variable: log(Lsp/L

up))

ATETTTUTPRTE: (1−α)TCMPRTE: Pα=P+α

MPRTE: Zkα=Zk+α

(b) Production Workers, Initial Non-Importers

8

C Estimating Hicks-Neutral Productivity

Our model implies that Hicks-neutral productivity differences are potentially among the most

important determinants of plant-level import decisions. Unfortunately, the data do not provide

a convenient measure of Hicks-neutral productivity. Moreover, standard productivity estima-

tion methods do not consider how we might separately identify skill-biased and Hicks-neutral

productivity.3 Accordingly, we develop an extension of the control function methods pioneered

by Olley and Pakes (1996) [OP, hereafter], Levinsohn and Petrin (2003) [LP, hereafter] and

Ackerberg, Caves and Frazer (2006), among others, to estimate a Hicks-neutral productivity

series for each plant in our data.4

We assume that the firm’s production function is specified as

Yit = eεitQit, where Qit = eα0+ωitKαkit M

αmit L

αpp,itL

αnn,it (13)

where ωit is the part of the Hicks-neutral productivity shock that is observed/anticipated by

firm i at the time which it makes input decisions while εit captures either measurement error or

an iid unanticipated shock that is not observed at the time which it makes input decisions. The

variables Lp,it and Ln,it represent the aggregate labor inputs for production and non-production

activities, respectively, and are defined by

Lj,it =

((AjL

sj,it)

σj−1

σj + (Luj,it)σj−1

σj

) σjσj−1

for j = p, n. (14)

Here, Lsj,it and Luj,it represent the number of skilled workers and that of unskilled workers,

respectively, in occupation j, where the subscript “p” indicates production workers while the

subscript “n” captures non-production workers. We assume that ωit follows a first order Markov

process.

To estimate the production function coefficients, including the elasticity of substitution pa-

rameters, we use the implications of plant profit maximization behavior.5 The first order con-

ditions with respect to Luj,it and Lsj,it are given by

W ut L

uj,it

Qit= αj

(Luj,it)σj−1

σj

(AjLsj,it)σj−1

σj + (Luj,it)σj−1

σj

andW st L

sj,it

Qit= αj

(AjLsj,it)

σj−1

σj

(AjLsj,it)σj−1

σj + (Luj,it)σj−1

σj

, (15)

3Doraszelski and Jaumandreu (2014) is a key exception.4Other important contributions to this literature include Wooldridge (2009), De Loecker (2011), De Loecker

et al. (2012) and Doraszelski and Jaumandreu (2014).5Our method is broadly based on the ideas contained in Gandhi, Navarro, and Rivers (2013), but our production

function is specified using a simple Cobb-Douglas form with CES aggregators for production and non-productionlabor inputs so that our analysis is substantially simpler than theirs.

9

respectively, so that (Luj,itLsj,it

) 1σ

A

σj−1

σj

j =W st

W ut

for j = p, n, (16)

where W st and W u

t represent the wages in year t for skilled and unskilled workers, respectively.

We assume that there is no unanticipated ex-post shock to Aj , Wst , and W u

t . Substituting (16)

into (14), we get

Lj,it = X−

σjσj−1

j,it Luj,it, where Xj,it ≡W ut L

uj,it

W st L

sj,it +W u

t Luj,it

.

Substituting the above equation for Lj,it into (13) and taking the logarithm gives

yit = α0,t + αkkit + αmmit + αplup,it + βpxp,it + αnl

un,it + βnxn,it + ωit + εit (17)

where βj = − σjαjσj−1 for j = p, n, and lower case letters represent the logarithm of the upper case

letters (e.g., yit ≡ ln(Yit)). Note that, if we can consistently estimate αj and βj , then we also

have a consistent estimate of σj because −βj/αj =σjσj−1 .

We recover the estimates in two stages. In the first stage, following LP, we write ωit as a

function of mit, kit: ωit = ω∗t (mit, kit). Taking an expectation of (17) conditional on (mit, kit),

and subtracting it from (17) gives

yit − E[yit|mit, kit] = αp{lup,it − E[lup,it|mit, kit]}+ βp{xp,it − E[xp,it|mit, kit]}

+αn{lun,it − E[lun,it|mit, kit]}+ βn{xn,it − E[xn,it|mit, kit]}+ εit.(18)

where E[εit|mit, kit] = 0 under the assumption that εit is mean zero random variable and that

εit is not observed yet when a plant makes intermediate input decision.

The parameters αp, βp, αn, and βp are estimated by (i) first estimating the functions

E[yit|mit, kit], E[`up,it|mit, kit], E[`un,it|mit, kit], E[xp,it|mit, kit] and E[xn,it|mit, kit] and then (ii)

running a no-intercept OLS regression of (18) using the estimate of the conditional expectation

terms. Note that, even though we consider the possibility of endogenous plant exit, the first

stage procedure is identical to that of LP.

In the second stage we identify the remaining production function parameters αk and αm.

To accomplish this, we first define

φt(mit, kit) ≡ α0,t + αkkit + αmmit + ω∗t (mit, kit)

and

xit ≡ yit − {αplup,it + βpxp,it + αnlun,it + βnxn,it}.

Further, let χit = 1 indicate plant survival in year t. We assume that a firm stays in the market

10

if and only if ωit ≥ ωt(kit) as in OP. Then, we may write (17) as

xit = α0,t + αkkit + αmmit + E[ωit|ωit−1, χit = 1] + ξit + εit

= αkkit + αmmit + gt(ωt(kit), ωit−1) + ξit + εit (19)

where ξit = ωit − E[ωit|ωit−1, χit = 1] and gt(ωt(kit), ωit−1) ≡ α0,t + E[ωit|ωit−1, χit = 1].

The survival probability conditional on ωt−1 is given by

Pr{χit = 1|ωit−1, kit−1,mit−1} = Pr{ωt ≥ ωt(kit)|ωit−1,mit−1, kit−1}

=

∫ ∞ωt(kit(mit−1,kit−1))

F (dωit|ω∗t−1(mit−1, kit−1))

= Pχit . (20)

where F (·) represents the law of motion for ωit. The capital stock follows kit = (1− δ)kit−1 + ιit

where ιit is the amount of investment between t − 1 and t, δ is the depreciation rate, and we

assume that ιit is a function of (ωit−1, kit−1) = (ω∗t (mit−1, kit−1), kit−1) so that we may write kit

as a function of mit−1 and kit−1, i.e., kit(mit−1, kit−1) in the second line of (20). We estimate the

survival probability (20) using a probit with third order polynomials in (mit−1, kit−1). Given

ω∗t−1(mit−1, kit−1), we may invert (20) with respect to ωt; therefore, we may write ωt as a

function of survival probabilities, Pχit , and ω∗t−1(mit−1, kit−1) as in ωt(Pχit , ω

∗t−1(mit−1, kit−1)).

Then, we may express gt(ωt(kit), ωit−1) in (19) as a (year-specific) nonlinear function of

(Pχit , ω∗t−1(mit−1, kit−1)) as

gt(ωt(Pχit , ω

∗t−1(mit−1, kit−1)), ω∗t−1(mit−1, kit−1))

= α0,t +

∫ ∞ωt(P

χit ,ω∗t−1(mit−1,kit−1))

ωitF (dωit|ω∗t−1(mit−1, kit−1))∫∞

ωt(Pχit ,ω∗t−1(mit−1,kit−1)) F (dωit|ω∗t−1(mit−1, kit−1))

.

Define

qt(Pχt , α0,t−1 + ω∗t−1(mit−1, kit−1)) ≡ gt(ωt(P

χit , ω

∗t−1(mit−1, kit−1)), ω∗t−1(mit−1, kit−1)),

and substituting this equation into (19) and using α0,t−1+ω∗t−1(mit−1, kit−1) = φt−1(mit−1, kit−1)−αkkit−1 − αmmit−1, we have

xit = αkkit + αmmit + qt(Pχt , hit−1) + ξit + εit, (21)

where hit = φt(mit, kit)− αkkit − αmmit. This equation corresponds to equation (12) in OP.

Given the above definitions, we recover αk and αm in three distinct steps. First, let xit =

yit − {αplup,it + βpxp,it + αnlun,it + βnxn,it}, where (αp, αn, βp, βn) are the first stage estimates of

the corresponding parameters. Then we estimate φ(mit, kit) by regressing xit on third order

11

polynomials in (mit, kit). Second, we estimate the survival probability by estimating the probit

for survival (χit = 1) conditional on (mit−1, kit−1) using third order polynomials. Third, for each

candidate value of (αk, αm), we compute hit(αk, αm) = φit − αkkit − αmmit and regress xit −{αkkit+αmmit} on third order polynomials in (Pχit , hit−1) to obtain the estimate of qt(P

χit , hit−1)

as its predicted value, denoted by qit(αk, αm). Denoting (ξit + εit)(αk, αm) = xit − {αkkit +

αmmit− qit(αk, αm)}, we estimate (αk, αm) using the moment conditions E[(ξit + εit)mit−1] = 0

and E[(ξit + εit)kit−1] = 0. Note that we do not use kit as an instrument because kit will be

correlated with ξit given that we take long differences.

We apply the above estimation procedure to the two years of data from 1996 and 2006 so that

the time subscripts t − 1 and t correspond to 1996 and 2006, respectively. The Hicks-neutral

productivity, including both the unexpected shock εit and the year-specific constant α0,t, is

computed as

ϕit ≡ α0,t + ωit + εit = yit − (αkkit + αmmit + αplup,it + βpxp,it + αnl

un,it + βnxn,it).

We find that (αk, αm, αp, αn, βp, βn) is estimated as (0.017, 0.602, 0.152, 0.110,−0.253,−0.138).

Note the production function parameters are very similar to those estimated elsewhere (e.g.

See Amiti and Konings (2007)). Our estimates further imply that the elasticity of substitu-

tion parameters among production and non-production workers (σp, σn) are estimated to be

(1.664,1.255).

As an alternative measure of productivity, we also estimate the “conventional” measure of

total factor productivity (TFP) under the assumption that skilled and unskilled workers are

perfect substitutes with a Cobb-Douglas production function given by

Yit = eεitQit, where Qit = eα0+ωitKαkit M

αmit L

αpp,itL

αnn,it (22)

where Lp,it = Lsp,it +Lup,it and Ln,it = Lsn,it +Lun,it. Repeating our estimation exercise under this

restriction we again recover the parameters (αk, αm, αp, αn) as (0.030, 0.908, 0.065, 0.074). We

also use this alternative structure and estimates to construct a second measure of productivity.

In the main text this second measure is denoted as “conventional” TFP.

D Capital-Skill Complementarity

As explained in Section 4.2.2, we consider the following capital-skill complementarity frame-

work. Output, Y , is produced using a production input, V p, and a non-production input,

V n, and intermediate materials, M : f(K,M,Ls, Lu, A, ϕ) = ϕ(V p)αp(V n)αnMαm with V j =

[(Aj(Ljs)β(Kj)1−β)1/ρj + (Lju)1/ρj ]ρj and ρj = σj/(σj − 1) for j = {n, p}. The firm’s cost min-

imization problem implies that we can write the following relationships between capital and

12

labor of each type:

Kp =

(Ws

Wk

)(1− ββ

)Lps and Kn =

(Ws

Wk

)(1− ββ

)Lns .

Therefore, total capital is related to total skilled labor as

K = Kn +Kp =

(Ws

Wk

)(1− ββ

)(Lps + Lns ), (23)

and it follows that the fraction of total capital allocated to occupation j ∈ {n, p} can be deter-

mined by dividing Kn or Kp by equation (23) as

Kj

K=

Ljs

Ljs + Lju.

Note that this result is sensitive to the assumption that the share of skilled labor and capital, β, is

equal across occupations. However, the alternative assumption that β varies across occupations

but capital is allocated in a fashion such that each firm has the same ratio of production to non-

production capital (i.e., Kn = γK and Kp = (1 − γ)K for some γ ∈ (0, 1)) results in a nearly

identical empirical structure. We do not find any significant difference using this alternative

assumption and, as such, we omit further discussion hereafter.

E First Stage Regressions

Tables E.1-E.2 present selected first stage regression results. We focus on the small instrument

set since the individual significance of the instrumental variables is more transparent.

13

Table E.1: Selected First Stage Regressions using a Linear Probability Model

Sample Full Initial Full Initial Full Full Full Initial Full Initial Full FullNon-Imp. Non-Imp. Non-Imp Non-Imp

Dep. Var. Import Import Import Import Import Import Standards Standards Standards Standards Import Capital-SkillDummy Dummy Dummy Dummy Dummy Dummy Dummy Dummy Dummy Dummy Dummy Ratio

Table-Column Table 4-(2) Table 4-(6) Table 5-(2) Table 5-(6) Table 7-(2) Table 7-(14) Table 8-(2) Table 8-(5) Table 8-(8) Table 8-(11) Table 9-(2) Table 9-(2)for 2nd StageTransport Cost -0.0734*** -0.0284*** -0.0714*** -0.268*** -0.0754*** -0.0729*** -0.0349*** -0.0365*** -0.0375*** -0.0399*** -0.0802*** 0.1854***

[0.009] [0.008] [0.009] [0.008] [0.009] [0.009] [0.013] [0.013] [0.013] [0.013] [0.011] [0.047]Import Air Share 0.3619*** 0.4380*** 0.3630*** 0.4419*** 0.3786*** 0.378*** -0.0123 -0.0264 0.0019 0.0003 0.4367*** 0.272

[0.131] [0.147] [0.130] [0.146] [0.131] [0.131] [0.153] [0.179] [0.153] [0.179] [0.149] [0.593]Import Weight 0.0050 -0.0006 0.0049 -0.0007 0.0050 0.0050 0.0051 0.0042 0.0051 0.0047 0.0032 0.0450

[0.010] [0.009] [0.010] [0.009] [0.010] [0.010] [0.011] [0.012] [0.011] [0.012] [0.012] [0.055]Export 0.1732*** 0.0821 0.1733*** 0.0823*** 0.1747*** 0.1753*** 0.1939*** -0.1347**

[0.015] [0.014] [0.015] [0.014] [0.015] [0.015] [0.016] [0.063]Capital 0.0257*** 0.0157*** 0.0253*** 0.0154*** 0.0271*** 0.0270*** 0.0342*** 0.0294*** 0.0346*** 0.0304***

[0.003] [0.003] [0.003] [0.003] [0.003] [0.003] [0.004] [0.004] [0.004] [0.004]Hicks-neutral ϕ 0.0158* 0.0127 0.015* 0.0125 0.0159* 0.0160* 0.0363*** 0.0141 0.0372*** 0.0167 0.0312*** 0.2371***

[0.009] [0.008] [0.009] [0.008] [0.009] [0.009] [0.011] [0.012] [0.011] [0.012] [0.010] [0.047]Foreign 0.1776*** 0.1207*** 0.1764*** 0.1189*** 0.1827*** 0.1812*** 0.0941 0.0891 0.1018* 0.0972 0.1560*** -0.1530

[0.032] [0.041] [0.032] [0.041] [0.032] [0.032] [0.059] [0.074] [0.058] [0.073] [0.035] [0.137]R&D 0.0628*** 0.0462** 0.0633*** 0.0464** 0.0653*** 0.0658*** 0.1161*** 0.1358*** 0.1183*** 0.1410*** 0.0675*** -0.0356

[0.021] [0.022] [0.021] [0.022] [0.021] [0.021] [0.035] [0.042] [0.035] [0.042] [0.022] [0.082]Training 0.0436*** 0.0290*** 0.427*** 0.0287*** 0.0478*** 0.0478*** 0.1241*** 0.1130*** 0.1267*** 0.1161*** 0.0613*** -0.0029

[0.011] [0.010] [0.012] [0.010] [0.011] [0.011] [0.015] [0.017] [0.015] [0.017] [0.013] [0.052]

ln(WsWu

) 0.0304 0.0199 0.0334 -0.0557 -0.0733* -0.0534 -0.0714* 0.0439 -0.227

[0.027] [0.021] [0.027] [0.037] [0.040] [0.037] [0.040] [0.034] [0.159]

ln(Lps

Lpu)96 0.0102** 0.0058 0.0146*** 0.0194 0.0162*** 0.0415**

[0.004] [0.004] [0.006] [0.006] [0.005] [0.020]

ln(LnsLnu

)96 0.0111** 0.0130 0.0069

[0.005] [0.008] [0.009]ln(Xp)96 -0.0705*** -0.0387*

[0.024] [0.022]ln(Xn)96 -0.0857***

[0.026]dpu,96 0.0714* 0.0667* 0.0394 0.0491 -0.0263 -0.0325 0.0842** 0.0332

[0.036] [0.039] [0.039] [0.042] [0.046] [0.051] [0.040] [0.148]dps,96 -0.0337*** 0.0026 -0.0677*** -0.0158 -0.0454*** -0.0437*** -0.0913*** -0.0303

[0.012] [0.010] [0.020] [0.018] [0.015] [0.016] [0.014] [0.072]dnu,96 -0.0200** -0.0477*** -0.0357*** -0.0271**

[0.009] [0.012] [0.010] [0.011]dns,96 0.0089 -0.0118 -0.0084 -0.0059

[0.010] [0.012] [0.012] [0.012](Cap-Skill)96 0.0185*** 0.2168***

[0.004] [0.022]No. Obs. 5,827 4,328 5,827 4,328 5,827 4,328 4,041 3,263 4,041 3,263 4,646 4,506R-squared 0.260 0.136 0.260 0.237 0.259 0.278 0.238 0.219 0.239 0.217 0.247 0.124

Notes: *** p<0.01, ** p<0.05, * p<0.1. Standard deviations are in parentheses. Province dummies and 3-digit ISIC industry dummies are also included.

14

Table E.2: Import Decision Model using Logit for the Sample of Production Workers

Full Sample Initial Non-ImportersCoeff. S.E. Ave. Deriv. S.E. Coeff. S.E. Ave. Deriv. S.E.

TC -0.5950 [0.1066] -0.0774 [0.0135] -0.3873 [0.1647] -0.0261 [0.0111]Air 0.1546 [0.0754] 0.2127 [0.1022] 0.3788 [0.1288] 0.3049 [0.1038]Wgt 0.0104 [0.0940] 0.0017 [0.0149] 0.0172 [0.1726] 0.0014 [0.0138]

TC × log(LpsLpu

)96 0.0764 [0.0799] 0.0049 [0.0050] 0.0220 [0.1285] 0.0008 [0.0044]

TC × dpu,96 0.0364 [0.0574] 0.0227 [0.0354] 0.0177 [0.2094] 0.0061 [0.0717]

TC × dps,96 -0.1043 [0.1237] -0.0225 [0.0263] -0.1288 [0.1483] -0.0142 [0.0165]

Air × log(LpsLpu

)96 -0.0195 [0.0701] -0.0112 [0.0395] 0.2266 [0.1224] 0.0736 [0.0398]

Air × dpu,96 0.0136 [0.0698] 0.0503 [0.2552] 0.0681 [0.1338] 0.1651 [0.3243]

Air × dps,96 0.0013 [0.0978] 0.0028 [0.2166] -0.0212 [0.1759] -0.0234 [0.1956]

Wgt× log(LpsLpu

)96 0.0522 [0.0877] 0.0036 [0.0060] 0.0793 [0.1733] 0.0028 [0.0061]

Wgt× dpu,96 0.0040 [0.0627] 0.0026 [0.0394] 0.0555 [0.1134] 0.0196 [0.0401]

Wgt× dps,96 0.1261 [0.1144] 0.0233 [0.0209] 0.0678 [0.1967] 0.0060 [0.0174]

Export 0.4723 [0.0462] 0.1196 [0.0108] 0.3941 [0.0704] 0.0552 [0.0100]Capital 0.4188 [0.0554] 0.0219 [0.0028] 0.4334 [0.0855] 0.0130 [0.0026]Hicks-neutral ϕ 0.0797 [0.0471] 0.0139 [0.0081] 0.1217 [0.0730] 0.0116 [0.0070]Foreign 0.1622 [0.0359] 0.0948 [0.0204] 0.1355 [0.0445] 0.0518 [0.0173]R&D 0.0836 [0.0386] 0.0315 [0.0142] 0.0957 [0.0496] 0.0213 [0.0111]Training 0.1735 [0.0504] 0.0386 [0.0111] 0.2125 [0.0792] 0.0251 [0.0093]

log(WsWu

)06 0.0365 [0.0574] 0.0229 [0.0356] 0.0436 [0.0966] 0.0137 [0.0305]

log(WsWu

)96 -0.0015 [0.0495] -0.0012 [0.0402] 0.0153 [0.0877] 0.0063 [0.0365]

log(LpsLpu

)96 0.0144 [0.1104] 0.0011 [0.0083] -0.1742 [0.2096] -0.0070 [0.0085]

dpu,96 0.0295 [0.1042] 0.0204 [0.0712] -0.0373 [0.2281] -0.0150 [0.0921]

dps,96 -0.2737 [0.1552] -0.0667 [0.0374] 0.0324 [0.2813] 0.0038 [0.0333]

No. Obs. 5716 4187

Notes: Estimates from the sample which uses the log of the production skill ratio as an outcome variable. Bootstrapstandard errors are in square brackets. Province dummies and 3-digit ISIC industry dummies are also included. Thesample excludes plants that belong to a 3-digit ISIC industry or province within which there is no variation in importstatus because, in such cases, the estimated coefficient of the corresponding industry or province dummy in the logitmodel would be either infinity or minus infinity.

15